perf(mpsc): streamline unbounded buffer reclamation

[tisonkun]

Summary

Unbounded receiver refills currently release retired batch allocations and oversized empty segment directories while holding the shared state mutex. Receiving also checks the last message and allocation capacity even for payload types whose segments cannot exceed the byte budget. Release retired storage after unlocking, and select oversized-payload reclamation at compile time so ordinary messages use the normal local pop path.

Make segment capacity and the retained directory's slot limit associated constants, with a named directory byte budget. Power-of-two growth keeps ordinary segments within 32 KiB; oversized payloads use one slot and release their backing allocation when consumed. A focused boundary test checks that growth, sealing, and buffer reuse preserve the capacity invariant around payload-size thresholds.

Validated with cargo x test on the default nightly and Rust 1.86.0, cargo x lint --fix, cargo x check, cargo x bench --no-run, and the full cargo x miri workflow. Benchmark adapters, extra measurement cases, and raw results are outside the repository.

Benchmarks compare main 7ad6d68d45a3b74506d9a605e594c417e80d506b with the final implementation in the same executable on an Apple M4 Max, macOS 26.6.2, rustc 1.99.0-nightly (3d6c19bb9). Values are medians of ten round medians, with randomized implementation order, 100 samples per case, 1,000 iterations per microbenchmark sample, and one batch per throughput sample. Coverage includes all 21 existing unbounded workloads, two 4,096-message inline bursts, and a 32-message burst of 64 KiB payloads. Async workloads transfer 16,384 messages per sample; inline async bursts use four runtime workers.

Workload	Main	This PR	Elapsed change
Try round trip	10.9350 ns	10.4600 ns	-4.3%
Sequential 64 B inline, 1,024 messages	10.4100 µs	8.1970 µs	-21.3%
Sequential 1 KiB inline, 1,024 messages	65.6400 µs	56.6300 µs	-13.7%
Sequential 1 KiB inline, 4,096 messages	265.1500 µs	226.6500 µs	-14.5%
Sequential boxed 1 KiB, 1,024 messages	55.0900 µs	53.2400 µs	-3.4%
Sequential 64 KiB inline, 32 messages	158.3000 µs	111.5000 µs	-29.6%
Async 1 KiB, 1 producer, 1,024-message bursts	1.7685 ms	1.3835 ms	-21.8%
Async 1 KiB, 8 producers, 1,024-message bursts	3.2125 ms	3.1105 ms	-3.2%
Native threads, 8 producers, 16,384 messages	360.7000 µs	360.5500 µs	-0.0%
Sequential usize, 1,024 messages, retained backlog	8.3330 µs	8.3740 µs	+0.5%

The sequential inline/boxed payload improvements occur in all ten paired rounds. Most usize throughput workloads remain close to main; the largest measured increase versus main is 0.5% in the 1,024-message retained-backlog case. These are workload-specific measurements on this machine, not a claim of uniform improvement across platforms.