Fix cancelled native StreamReader waiters and preserve buffered data

[egorsmkv]

Fixes #106.

A cancelled native stream read leaves its Future registered in ReadWaiter. Retrying before data arrives raises the concurrent-reader ValueError; later feeds can instead consume and discard bytes for the cancelled Future. Fragmented readexactly() additionally owns bytes in a private accumulator.

This change:

• releases cancelled waiters through a weak-reference done callback plus cancellation checks before reads and data/EOF/error processing;
• restores only the initialized exact-read prefix ahead of newer buffered data and reapplies flow control;
• checks Future identity so an old queued callback cannot clear a newer waiter;
• preserves active concurrent-read rejection and the immediate buffered-read path.

Adds differential cancellation/timeout/retry/EOF/error tests on asyncio, rsloop and uvloop; a real private Redis server integration test for redis-py 5.3.1 + hiredis (SET/GET, both pipeline modes, connection reuse and PING); a Linux CI job; a reproducible benchmark. Updates an existing regression that incorrectly expected cancelled exact reads to lose bytes.

Actual local validation

Linux x86_64, CPython 3.14.7, nightly-2026-09-25, default-feature release extension, Redis 7.2.5, redis-py 5.3.1, hiredis 3.4.2, uvloop 0.23.0.

• Source baseline f16e637: zero-timeout retry, TCP reproduction and Redis SET all fail with the reported ValueError. asyncio/uvloop reference scenarios pass.
• Targeted cancellation + Redis + existing stream tests: 226 passed.
• Full pytest: 453 passed, 3 skipped, 2 failed, 98 deselected. The two Unix-socket failures are EPERM and also reproduce with the baseline extension.
• Rust default: 319 passed, 15 failed. All failures are io_uring driver creation rejected with EPERM.
• Rust all features: 415 passed, 19 failed, likewise blocked io_uring construction (including four filesystem tests).
• Clippy all targets/all features with -D warnings, rustfmt, Ruff and focused Pyright passed.

The full suites did not pass in this environment. Network tests exercised mio/epoll fallback. Actual io_uring, macOS, Windows, other Python versions and free-threaded Python were not validated locally; CI remains a separate gate.

Performance tradeoff

Same-toolchain release builds, baseline/fix/fix/baseline, CPU-pinned lifecycle benchmark:

• Ready buffered reads show no observed slowdown; small differences are not claimed as improvements.
• Pending read rounds cost +35–50% in the synthetic benchmark. Exact/until reads add approximately 0.43–0.50 µs per round; read() uses two pending reads per round and adds 0.84 µs. This is measurable cancellation-check/callback overhead, not a performance optimization.
• TCP benchmark results changed direction across repetitions; no reliable end-to-end regression or improvement is established.

CI follow-up (2026-10-10)

Merged master at 439d288, preserving the existing PR commits and file removals.

• Added the 3 missing cancellation profiling hooks and 7 hooks for the reactor/STARTTLS code brought in from master. The existing audit now reports 2076 function definitions audited; 0 missing hooks. No exclusions or coverage thresholds were relaxed.
• Native test output now goes to regular log files before being printed, avoiding the Windows/MSYS pipe failure seen in run 94's predecessor (the failure pattern is documented in rust-lang/rust#98947). All-feature and isolated-feature tests still preserve nonzero exit status, full logs and artifacts; no tests were removed.
• Restricted the private reader._transport identity assertion to rsloop. Older asyncio keeps a different internal transport; every backend still checks encrypted round trips, writer replacement, EOF, callback count and shutdown. Reproduced both failing reference cases locally on CPython 3.12.3 before this adjustment; both pass afterward.

Local follow-up validation:

• Release builds with profile and hotpath-alloc-profile succeeded.
• Each build: 252 passed, 3 skipped, 1 deselected for profiler/stream/cancellation/TLS tests. Redis tests were skipped because redis-server is unavailable on PATH; the known restricted Unix-socket case was deselected.
• Cancellation profiling produced a report in both modes.
• Coverage-parser Rust tests: 3 passed. Profiling tooling tests: 15 passed.
• All-target/all-feature Clippy with warnings denied, rustfmt, Ruff, focused Pyright and diff whitespace checks passed.
• Checked both modified shell loops with successful and failing commands: stdout/stderr remain in the logs and exit code 23 remains a failure.

On GitHub Actions, run 94 passed the standard profiling job, Windows native/isolated-feature job, both Rust suites and Redis integration. It exposed the older-asyncio assertion fixed in the final commit. The allocation-profile job was cancelled when that commit started run 95; final CI results are pending.