perf(spmc): specialize single-producer coordination

[tisonkun]

Summary

SPMC used a producer waiter queue even though send(&mut self) permits only one pending send. Replace that queue with one waker slot, and carry the exclusive producer capability into the implementation. Add cancellation, notification-race, reentrant-callback, and payload-destruction tests, plus SPMC primitive benchmarks.

Validation: cargo x test, cargo x check, cargo x lint, and cargo x miri. All 28 SPMC tests pass normally; Miri passes 26, with the two existing Tokio runtime tests excluded.

Design Notes

• The bounded sender owns its capacity limit. Consumers only remove values: once they free space, no competing producer can claim it. Sending needs no waiter ID, node allocation, capacity grant, or cancellation handoff.
• Keep the immediate try_send fast path. Pending sends reuse the single waker slot; receiving checks the queue and registers interest in one critical section. Unbounded publication has no capacity check or Full path.
• Keep the shared mutex and receiver waiter queue: consumers still compete for values and must transfer unconsumed notifications on cancellation. Wake callbacks and waker/payload destruction remain outside the lock. No new unsafe code is introduced.
• Bounded storage is preallocated, moving allocation to channel creation instead of growing the buffer while sending. This requires the full requested buffer allocation upfront; the documentation and changelog record the change.

Local measurements on Apple M4 Max, rustc 1.99.0-nightly (3d6c19bb9), against f998f35 with the same benchmark cases. Values are medians of five run medians, alternating old/new execution order.

Primitive cycle	Before	After	Time change
Cancel pending send	19.92 ns	16.67 ns	-16.3%
Wake blocked sender	29.69 ns	24.65 ns	-17.0%
Wake receiver, bounded	32.46 ns	24.64 ns	-24.1%
Wake receiver, unbounded	32.94 ns	24.97 ns	-24.2%
Ready send/receive, bounded	14.88 ns	15.37 ns	+3.3%

For the existing 16,384-message task batch with four runtime workers:

Consumers	Bounded before → after	Unbounded before → after
1	391.4 → 397.3 µs (+1.5%)	235.0 → 252.9 µs (+7.6%)
2	748.3 → 652.8 µs (-12.8%)	538.0 → 435.6 µs (-19.0%)
4	925.2 → 821.8 µs (-11.2%)	827.6 → 749.0 µs (-9.5%)
8	1064 → 981.4 µs (-7.8%)	817.0 → 779.5 µs (-4.6%)

The gains are concentrated in waiting and competing-consumer workloads; this is not a universal throughput improvement. In particular, the unbounded single-consumer task batch regressed in these local measurements. Current-thread task-batch changes ranged from -4.7% to +2.0%.

cargo x bench --bench primitives -- 'spmc::' --sample-count 2000
cargo x bench --bench ecosystem -- 'spmc::.*Spmc' --sample-count 100