The PMU-slot analysis in the type checker hard-coded the perf_event bit
layout (value >> 32, type tags 1/2) and the enum name "perf_event",
duplicating the packing that stdlib already defines in the enum values.

Move the layout to stdlib: the perf_event enum is now built from
perf_event_pack, and perf_event_consumes_pmu_slot decodes with the same
named type constants. The type checker no longer knows the bit layout or
the enum name — it resolves the event constant generically across enum
definitions and asks stdlib whether it consumes a PMU slot. Adding a new
event to the stdlib enum now classifies correctly with no type_checker
change.

Signed-off-by: Cong Wang <cwang@multikernel.io>

The type checker rejected perf event groups that needed more PMU counter
slots than a limit read from the *build* host's sysfs (or an env var, or
a guessed default of 4). This was the wrong layer: it's a host-dependent
resource check, not a type rule, and the build host may not match the
deploy host. The kernel already enforces group sizing authoritatively at
perf_event_open(2) on the real host, and the generated ks_open_perf_event
reports that failure at runtime.

Deleting it removes ~240 lines and all the perf-group/PMU/event-encoding
knowledge from type_checker.ml — perf_type references drop from 55 to 11,
and those 11 are the same uniform per-program-type handling @xdp/@kprobe
get (attribute->program-type, signature check, attach() overload). The
type checker no longer understands perf groups at all.

Also drops the now-orphaned stdlib helpers (perf_event_consumes_pmu_slot,
perf_type_tracepoint) and the three tests that asserted the static check.
Docs updated to describe runtime/kernel enforcement.

Signed-off-by: Cong Wang <cwang@multikernel.io>

perf_options exposed two ways to join a perf group: the typed `group`
(a PerfAttachment handle) and the raw `group_fd: cache.perf_fd` form,
which leaked a kernel ABI integer into the user-facing struct. In
KernelScript every perf fd comes from attach() as a PerfAttachment, so
the raw form was redundant.

Remove the public `group_fd` field (and its default). Grouping is now
expressed only via `group`. Internally, ks_open_perf_event still resolves
a leader fd from the handle and defaults to -1 (no group); the generated
runtime keeps its group_fd plumbing. Passing `group_fd:` now fails type
checking as an unknown field.

Tests, examples, and docs updated; the test for the raw group_fd form is
removed and its detach-cascade assertions folded into the group-handle
test.

Signed-off-by: Cong Wang <cwang@multikernel.io>

The three user-facing perf types were inconsistently cased: perf_options
(snake) alongside PerfAttachment and PerfRead (Pascal). Rename the latter
two to perf_attachment and perf_read so the whole perf surface is
snake_case, matching the rest of the language's builtin types.

KernelScript type name and generated C type name stay identical (the
generic IRStruct -> "struct <name>" mapping handles it), so no codegen
special-casing is needed; neither name collides with kernel structs.
Examples, tests, and docs updated.

Signed-off-by: Cong Wang <cwang@multikernel.io>