perf(completion): make terminal waits lock-free

[ryux1]

Closes #246.

Summary

• Publish completion status through an atomic so completed and abandoned waits avoid the waiter mutex.
• Keep pending registration cancellable with one lock and a status recheck, while cloning, dropping, and waking wakers outside the lock.
• Add a regression test where RawWaker::clone drops the completer, proving clone callbacks cannot deadlock against completion state.

Design notes

Completion and abandonment hold the waiter mutex while detaching the current waiter cohort, then publish the terminal state with a release store. Polls use acquire loads, so observing Completed also observes the initialized OnceLock value and observing either terminal state happens after waiter detachment. A poll that initially sees Pending clones its waker before locking, then rechecks status under the waiter mutex before registering. Wait::drop uses the same load/lock/recheck pattern.

Three cargo x bench runs were collected for upstream main (93b3413) and this branch on an Intel Core i9-10900F (Linux x86_64); the table reports the median of each set:

Completion benchmark	main	this branch	change
ready_wait	16.43 ns	6.848 ns	-58.3%
cancel_pending	36.91 ns	40.04 ns	+8.5%
complete_then_wait	86.88 ns	78.16 ns	-10.0%
notify_pending	121.9 ns	113.9 ns	-6.6%
fanout/1	144.1 ns	134.9 ns	-6.4%
fanout/2	199.1 ns	183.2 ns	-8.0%
fanout/4	308.0 ns	276.5 ns	-10.2%
fanout/8	577.1 ns	531.0 ns	-8.0%
fanout/32	2.573 us	2.380 us	-7.5%

The pending cancellation path pays about 3.1 ns to clone outside the lock; terminal waits and every measured completion/notification path improve.

Validated with:

• cargo x check
• cargo x test
• cargo x lint
• cargo x miri
• cargo +nightly miri test --package tests-integration --test completion_test
• cargo x bench