perf(semaphore): store wake batch on stack

[tisonkun]

Summary

• replace the lazily allocated Vec<Waker> in semaphore release with a fixed 32-slot stack batch
• preserve the existing unlock-before-wake batching and FIFO wake order
• keep panic cleanup explicit: start..end is the only initialized range, and advancing start before wake lets unwind drop the remaining wakers

This is stacked on #189 so the Waker correctness changes remain independently reviewable.

A shorter [Option<Waker>; 32] prototype regressed the empty-queue release benchmark by about 18%. The final MaybeUninit helper avoids initializing unused slots and keeps a single documented invariant.

Benchmarks

Apple Silicon, medians, alternating this branch with #189 using 0.5s minimum time and 80 samples:

Benchmark	#189	This PR	Change
handoff_permit	124.4 ns	82.55 ns	-33.6%
queued_owned_burst/1	249.4 ns	207.5 ns	-16.8%
queued_owned_burst/8	665.4 ns	556.8 ns	-16.3%
queued_owned_burst/32	2.290 us	1.812 us	-20.9%
release	22.11 ns	20.71 ns	-6.3%

Command:

cargo bench -p benchmarks --bench benchmarks -- semaphore::handoff_permit semaphore::queued_owned_burst semaphore::release --min-time 0.5 --sample-count 80 --color never

Testing

• cargo +1.86.0 check -p asyncband --all-features
• cargo x lint
• cargo x test --no-capture