Fix cancelled timer payload lifetimes and compact timer queues

[egorsmkv]

Fixes #118.

Cancelled timers behind an earlier live deadline kept callback, args and context until heap removal, pinning completed timeout tasks and closed asyncpg statement caches. TimerCallback now takes Python ownership under a timer-only mutex and releases it outside the lock. Dispatch claims ownership under the same lock and keeps running callbacks valid through concurrent cancellation. The existing call_soon/Handle layout and invocation path are unchanged.

Both timer queues compact after more than 100 entries and over half the queue are cancelled. Idempotent accounting handles cross-thread cancellation/removal and shutdown; removed shells go to the ready drain. Tests cover immediate ownership release, contexts, repeated cancellation, finalizer reentry, cancellation after dequeue, expired timeouts, running invocation ownership and heap ordering.

TDD: the first commit fails 12 lifetime regressions on 0.1.63; a local Rust regression found 257 queued entries instead of 1. Real baseline PostgreSQL CI retains 500 caches and 1500 prepared statements after 500 closed connections; asyncio/uvloop retain 0. Fixed PostgreSQL 3.14 CI reports 0 for caches, statements and protocols, with 39 lifecycle tests passing. Fixed local 20,000-timer/timeout probes retain 0 payloads/tasks after turns: Python retention drops from 87.2 MB / 101.0 MB to 307 B / 29.9 KB. Timer RSS drops from 125.2 to 33.1 MiB and timeout RSS from 150.4 to 34.1 MiB in fresh-process probes.

Validation: 51 differential tests passed (6 intentional control skips); 19 new free-threaded tests plus 5 existing timer tests passed. Local core: 628 passed; ecosystem: 78 passed. Two core Unix-socket failures reproduce on the baseline; local Rust syscall-restricted cases are documented, not counted as passing. Clippy with all targets/features, fmt, Ruff and targeted Pyright passed. CI is being monitored; a Python 3.10-only benchmark compatibility failure was corrected to use wait_for and skip unavailable asyncio.timeout.

Performance: two reversed-order release-build pairs, 14 samples per build, 1M operations/sample. call_soon is stable (-0.4%, noise); timer-only synchronization costs about 15–30 ns/op (+5.8% retained/cancelled, +11.5% discarded). Scattered cancelled timers measured -2.7%; no general speedup claim.

Measurements and reproduction probes are in benches/timer_cancellation.py and the baseline PostgreSQL CI / passing fixed CI. The report file was removed before merge.

The PostgreSQL/memory checks run against a real CI PostgreSQL service. The author's 300-second Granian production workload was not repeated.