perf(mpsc): coalesce unbounded receiver wakeups

[tisonkun]

Summary

• avoid shared write operations on every unbounded send when no receive call is parked
• retain the standard-library queue core and the existing public API without adding dependencies
• add deterministic coverage for burst notification coalescing and last-sender disconnect

A local Apple Silicon comparison used the existing ecosystem benchmark with two seconds per benchmark. Median latency changed as follows:

Case	main	This PR	Change
ready round trip	9.185 ns	7.340 ns	-20.1%
try round trip	8.290 ns	6.566 ns	-20.8%
concurrent, 1 producer	177.9 us	148.6 us	-16.5%
concurrent, 2 producers	1.003 ms	871 us	-13.2%
concurrent, 4 producers	2.124 ms	1.789 ms	-15.8%
concurrent, 8 producers	3.498 ms	3.551 ms	+1.5%

The eight-producer result is effectively unchanged; contention in the standard-library MPMC queue remains the limiting factor there.

Design Notes

The queue remains the durable source of truth. The receiver follows register waker -> publish waiting -> recheck queue, while a successful sender follows enqueue -> claim waiting -> wake. The waiting flag uses sequentially consistent operations.

If a sender claims the flag, it owns the wake. If it observes the pre-waiting false value, its enqueue precedes the receiver's subsequent queue recheck, so the receiver observes the message instead of sleeping. Clearing the flag after a successful recheck prevents stale notification ownership. Last-sender disconnect uses the same notification path after dropping the standard-library sender.

This keeps the proof local to notification delivery, adds no unsafe code or dependency, and intentionally leaves replacement of the queue backend in #209 out of scope.

Validated with cargo x test, cargo x check, and cargo x lint.