perf(mpsc): replace boxed receiver wakers

[tisonkun]

Summary

• replace the heap-allocated AtomicOptionBox<Waker> used by bounded and unbounded MPSC receivers with an internal AtomicWaker
• keep the MPSC feature dependency-free and inline the registration and wake-up hot paths
• preserve notification ordering across register/wake races and recover cleanly from panicking custom wakers
• benchmark repeated receiver registration and a complete pending/wake/receive cycle

Rationale

Every pending receiver poll currently clones a Waker, allocates a Box, atomically exchanges its pointer, and drops the previous allocation. The new state machine stores the waker inline and skips cloning when the registered waker already targets the same task.

The existing check-register-check sequence remains unchanged, so a wake that happens before registration is still observed through the second queue check. A wake that races with registration is completed by whichever side owns the inline waker slot.

The implementation remains private to Asyncband instead of adding an atomic-waker-style dependency. On 64-bit targets this trades approximately 16 additional inline bytes per channel state for eliminating allocations on pending polls.

Benchmarks

Local median results from repeated runs of cargo bench -p benchmarks --bench benchmarks -- mpsc::reregister_pending_receiver mpsc::wake_pending_receiver:

Benchmark	Before	After	Change
reregister_pending_receiver	20.3–21.2 ns	5.8–6.3 ns	3.3–3.6x faster
wake_pending_receiver	31.3–32.0 ns	16.7–21.5 ns	1.5–1.9x faster

Validation

• cargo +nightly x lint
• cargo +nightly x check
• cargo +1.86.0 x test
• cargo test --workspace --all-features