refactor: simplify unbounded MPSC and consolidate validation

[tisonkun]

Summary

• Keep unbounded MPSC endpoint state under one mutex and replace receiver waker registration retries with a linear path. Preserve receiver-local batches, FIFO, cancellation, and incremental storage reclamation while removing the spare-segment cache; organize endpoint definitions with their implementations and clarify delivery contracts.
• Consolidate shared integration-test wakers, polling, and deadlock checks. Establish pending waiters deterministically, verify complete MPSC message sets, remove subsumed tests, and keep necessary reclamation checks next to their state.
• Remove avoidable benchmark adapter allocation and incomplete worker round trips. Consolidate duplicate suites and reduce parameter matrices while retaining representative scheduling, backpressure, payload, and backlog scenarios.
• Add cargo x bench --no-run and compile benchmarks in the existing required check job. Local cargo x bench continues to run measurements on demand.

Validation: cargo x test (485 passed), cargo x check, cargo x lint, cargo x miri (89 passed; three OS-runtime cases skipped by the existing configuration), cargo x bench --no-run, and both benchmark binaries with --test to run every retained case once. Independent test and benchmark reviews found no remaining actionable issues.

Design Notes

The receiver still consumes its local batch without locking. Sender clone/drop now uses the shared mutex: local M4 Max measurements show about 5.5 ns additional cost per clone/drop pair and 1 ns per ready send/receive pair. Scheduled scalar batches were roughly unchanged; eight OS-thread producers with a blocking receiver were 12% slower, while eight-producer inline 1 KiB bursts were 13% faster. These are five-round local medians across 16 representative cases, with concurrent scheduling noise, rather than performance guarantees.