There's no reason a git repository couldn't be at the root of a server,
and URLs should have an implicit path of '/' when one is not specified.

We did not properly support default credentials for proxies, only for
destination servers.  Refactor the credential handling to support sending
either username/password _or_ default credentials to either the proxy or
the destination server.

This actually shares the authentication logic between proxy servers and
destination servers.  Due to copy/pasta drift over time, they had
diverged.  Now they share a common logic which is: first, use
credentials specified in the URL (if there were any), treating empty
username and password (ie, "http://:@foo.com/") as default credentials,
for compatibility with git.  Next, call the credential callbacks.
Finally, fallback to WinHTTP compatibility layers using built-in
authentication like we always have.

Allowing default credentials for proxies requires moving the security
level downgrade into the credential setting routines themselves.
We will update our security level to "high" by default which means that
we will never send default credentials without prompting.  (A lower
setting, like the WinHTTP default of "medium" would allow WinHTTP to
handle credentials for us, despite what a user may have requested with
their structures.)  Now we start with "high" and downgrade to "low" only
after a user has explicitly requested default credentials.

Update our CI tests to start a proxy that requires NTLM authentication;
ensure that our WIndows HTTP client can speak NTLM.

Increase the permissible replay count; with multiple-step authentication
schemes (NTLM, Negotiate), proxy authentication and redirects, we need
to be mindful of the number of steps it takes to get connected.

7 seems high but can be exhausted quickly with just a single authentication
failure over a redirected multi-state authentication pipeline.

We cannot examine the keep-alive status of the http parser in
`http_connect`; it's too late and the critical information about whether
keep-alive is supported has been destroyed.

Per the documentation for `http_should_keep_alive`:

> If http_should_keep_alive() in the on_headers_complete or
> on_message_complete callback returns 0, then this should be
> the last message on the connection.

Query then and set the state.

Some authentication mechanisms (like HTTP Basic and Digest) have a
one-step mechanism to create credentials, but there are more complex
mechanisms like NTLM and Negotiate that require challenge/response after
negotiation, requiring several round-trips.  Add an `is_complete`
function to know when they have round-tripped enough to be a single
authentication and should now either have succeeded or failed to
authenticate.

When we get an authentication failure, we must consume the entire body
of the response.  If we only read half of the body (on the assumption
that we can ignore the rest) then we will never complete the parsing of
the message.  This means that we will never set the complete flag, and
our replay must actually tear down the connection and try again.

This is particularly problematic for stateful authentication mechanisms
(SPNEGO, NTLM) that require that we keep the connection alive.

Note that the prior code is only a problem when the 401 that we are
parsing is too large to be read in a single chunked read from the http
parser.

But now we will continue to invoke the http parser until we've got a
complete message in the authentication failed scenario.  Note that we
need not do anything with the message, so when we get an authentication
failed, we'll stop adding data to our buffer, we'll simply loop in the
parser and let it advance its internal state.

We must always consume the full parser body if we're going to
keep-alive.  So in the authentication failure case, continue advancing
the http message parser until it's complete, then we can retry the
connection.

Not doing so would mean that we have to tear the connection down and
start over.  Advancing through fully (even though we don't use the data)
will ensure that we can retry a connection with keep-alive.

Ensure that the server supports the particular credential type that
we're specifying.  Previously we considered credential types as an
input to an auth mechanism - since the HTTP transport only supported
default credentials (via negotiate) and username/password credentials
(via basic), this worked.  However, if we are to add another mechanism
that uses username/password credentials, we'll need to be careful to
identify the types that are accepted.

Include https://github.com/ethomson/ntlmclient as a dependency.

A "connection" to a server is transient, and we may reconnect to a
server in the midst of authentication failures (if the remote indicates
that we should, via `Connection: close`) or in a redirect.

Hold an individual authentication context instead of trying to maintain
all the contexts; we can select the preferred context during the initial
negotiation.

Subsequent authentication steps will re-use the chosen authentication
(until such time as it's rejected) instead of trying to manage multiple
contexts when all but one will never be used (since we can only
authenticate with a single mechanism at a time.)

Also, when we're given a 401 or 407 in the middle of challenge/response
handling, short-circuit immediately without incrementing the retry
count.  The multi-step authentication is expected, and not a "retry" and
should not be penalized as such.

This means that we don't need to keep the contexts around and ensures
that we do not unnecessarily fail for too many retries when we have
challenge/response auth on a proxy and a server and potentially
redirects in play as well.

For request-based authentication mechanisms (Basic, Digest) we should
keep the authentication context alive across socket connections, since
the authentication headers must be transmitted with every request.

However, we should continue to remove authentication contexts for
mechanisms with connection affinity (NTLM, Negotiate) since we need to
reauthenticate for every socket connection.

Instead of using `is_complete` to decide whether we have connection or
request affinity for authentication mechanisms, set a boolean on the
mechanism definition itself.

We stop the read loop when we have read all the data.  We should also
consider the server's feelings.

If the server hangs up on us, we need to stop our read loop.  Otherwise,
we'll try to read from the server - and fail - ad infinitum.

When we have a keep-alive connection to the server, that server may
legally drop the connection for any reason once a successful request and
response has occurred.  It's common for servers to drop the connection
after some amount of time or number of requests have occurred.

When we're issuing a CONNECT to a proxy, we expect to keep-alive to the
proxy.  However, during authentication negotiations, the proxy may close
the connection.  Reconnect if the server closes the connection.

When we send HTTP credentials but the server rejects them, tear down the
authentication context so that we can start fresh.  To maintain this
state, additionally move all of the authentication handling into
`on_auth_required`.