coro: demand-gate native syscall compensation

[cpunion]

Moves compiler-certified llgo.syscall calls off the generic worker transaction and onto one same-M entersyscall/exitsyscall boundary, while preserving worker/locked-M behavior for ordinary C and cgo calls.

Highlights:

• structurally omit worker park/resume CFG and worker state from certified native syscall lowering
• classify replacement demand at the detached scheduler boundary with one common external-source count plus the orthogonal TaskControl endpoint count
• asynchronously request/cancel cached clean M owners with exact dispatch/claim/return races
• fail closed on ambiguous managed-quota release and correctly follow replacement retirement successors by generation baton
• add direct-syscall and genuinely blocking pipe fixtures, design gates, and measured performance checkpoint

Validation:

• full runtime go test -p=1 -timeout=180s ./...
• focused compiler syscall/worker tests
• internal/build complete TestCoro* suite
• native corofleet C harness at -count=20
• replacement-owner retirement E2E at -count=10
• fresh LLVM 22 linked file, direct-syscall, TCP, and sole-M blocking-pipe binaries
• 500 direct-file, 200 standard-file, 200 TCP, and 100 blocking-pipe independent-process stress run

31-run AB/BA medians for 500 operations versus this PR parent:

• direct syscall file: 5.379 ms -> 2.240 ms (-58.4%)
• standard os.File: 8.772 ms -> 5.556 ms (-36.7%)
• TCP: 20.454 ms -> 20.372 ms (neutral)

The remaining source-free direct-channel cross-route dependency is documented explicitly: its completed producer request should activate a deferred replacement, rather than treating every unmatched channel waiter as eager syscall demand.