Add bounded timer waits and cancellation callbacks

[renerocksai]

Return-and-resume handlers can now wait on timers without keeping worker threads occupied. Opt-in cancellation callbacks release retained application state after abandoned flushes and waits.

• Add Context.wait(delay_ns), .timer, and opt-in .cancelled callbacks.
• Retain request deadlines, executor ownership, and transport borrows through cancellation.
• Drain earlier pipeline responses before a subsequent request waits.
• Require local cleanup before .finish or .close; those actions never trigger an extra callback.
• Keep timer storage bounded per slot, without new threads, kernel operations, or request-time allocation.

The timer uses a fresh monotonic timestamp. Existing polling can delay wakeups. This interface does not provide dynamic output leases. Application state and locals remain ineligible for output borrowing.

Exact candidate 6f6af03e828159455b91cece6ba468201651e1f6 passed native verification with Zig 0.16.0. Debug and ReleaseSafe each completed all 16 build steps.

Platform	Test executions per mode	Existing wire groups	Continuation groups
Linux x64	109 passed, 2 platform skips	84 passed	9 passed
macOS ARM64	107 passed, 4 platform skips	84 passed	8 passed; unsupported three-owner topology skipped
Windows x64	122 passed, 4 platform skips	80 passed	9 passed

Linux/macOS native run and Windows native run passed. Windows also passed nine shard cases and the 30,000-response correctness smoke gate. Documentation and offline whitepaper checks passed locally.

Fixtures cover 32 waiters sharing one worker, depth-128 pipelines, partial sends, one-slot reuse, and shutdown. Deterministic tests cover terminal-result races, queued cancellation, stale clocks, and frozen batches. The first Windows attempt exposed an omitted socket-error class in the test retry; the final candidate includes that correction. No Zig source changed for that correction. No performance tests were run.