NTS-surfaced test failures

[EdmondDantes]

Tracking issue for three test failures surfaced once the NTS build was fixed
(commit a167a92 — before that, NTS jobs failed at build and never ran tests,
so these were latent). None are regressions from the spawn_thread shutdown
fix; all are pre-existing.

Run that exposed them: https://github.com/true-async/php-async/actions/runs/25852397574

1. cancel_during_io — getaddrinfo struct leak (NTS)

fuzzy-tests/_generated/cross_topic/cancel_during_io__00_cancel_a_coroutine_blocked_on_tcp_accept.phpt
fails on LINUX_X64_DEBUG_NTS with:

libuv_reactor.c(3010) : Freeing 0x... (288 bytes) ...
=== Total 1 memory leaks detected ===

This is the second half of #111. da62db4 fixed the leak of the libuv
addrinfo result buffer (event->result). The 288-byte block at
libuv_reactor.c:3010 is the pecalloc of the async_dns_addrinfo_t
event struct itself, which still leaks: in libuv_dns_getaddrinfo_dispose,
the !LIBUV_DNS_F_CALLBACK_DONE path sets LIBUV_DNS_F_DISPOSE_PENDING,
uv_cancels and returns without pefree. It relies on on_addrinfo_event
firing again to re-enter dispose and free the struct — but if the reactor is
torn down before that libuv callback fires, the struct leaks.

2. thread_pool/cancel — flake under tracing JIT (ZTS)

fuzzy-tests/_generated/thread_pool/cancel__00_submit_then_cancel_every_future_settles_cleanly.phpt
failed once on LINUX_X64_RELEASE_ZTS under the "Test Tracing JIT" step.
Passes 5/5 locally (ZTS debug + tracing JIT). Suspected flake; needs
repeated runs to characterise.

3. curl progress-callback exception (macOS)

tests/curl/035-progress_exception.phpt (and 056-multi_progress_exception.phpt)
fail on MACOS_ARM64_*_NTS with Fatal error: Uncaught RuntimeException: Progress callback error — the script dies before Done. Also failed on
macOS ZTS in the very first run, so it is macOS-specific and pre-existing.
An exception thrown from the CURLOPT_XFERINFOFUNCTION callback is not being
delivered to the awaiter on macOS.

Iteration workflow

.github/workflows/debug-bugfix.yml — manually-triggered (workflow_dispatch),
builds only the platform/build combos that reproduce each failure (Linux NTS

• ASAN, Linux ZTS + tracing JIT, macOS NTS) and runs just these four tests
  with --repeat 5. Trigger with:

gh workflow run "Debug Bugfix (manual)" --repo true-async/php-async --ref main

[EdmondDantes]

Confirmed pre-existing — not regressions. Commit d1be46b (adds only the
debug-bugfix.yml workflow file, zero code changes — byte-identical code
to a167a92) still fails CI on curl/035-progress_exception and
056-multi_progress_exception on macOS (NTS and ZTS). So #3 is a
deterministic pre-existing macOS bug.

The NTS build fix only made the NTS jobs run (they previously failed at the
build step and never reached the test phase), which surfaced these latent
failures.

Status from the first Debug Bugfix run + local repro:

• Implement the functionality of the Scope class. #2 cancel_during_io leak — did NOT reproduce under Linux NTS + ASAN
  (×5), nor locally (×many). Very environment/timing specific.
• Whether Coroutine\Channel and Coroutine\Defer are supported #1 thread_pool/cancel — reproduced as a segfault (Termsig=11)
  under RELEASE ZTS + tracing JIT, ~0.5% rate; not catchable under gdb (race
  masked) and not caught in 480 local parallel runs. thread_pool already had
  documented pre-existing shutdown races.
• Fix test execution to pass on both Unix and Windows OS #3 curl/035+056 — deterministic on macOS, see above.

debug-bugfix.yml updated: macOS bison PATH fix, --repeat raised to 30.