perf: reduce fan-out notification overhead

[tisonkun]

Summary

• Separate reusable waiter draining from terminal take_all, moving terminal arena scanning and destruction outside the state lock.
• Share an inline-one WakerBatch across wait-set, event, condition-variable, and semaphore fan-out paths without adding a dependency.
• Keep reusable paths retaining their arena allocation and use take_all only for terminal completion or disconnection transitions.

Design Notes

• Interleaved Completion benchmarks show 10.5%–13.0% lower end-to-end latency for 1–8 waiters and 4.2% for 32 waiters.
• An inline-eight SmallVec prototype improved larger fan-out but did not justify a new dependency or a bespoke unsafe collection without production fan-out evidence, so this change keeps the focused inline-one representation.

Validation

• cargo x check
• cargo x test
• cargo x lint