Pause a stopped listen event instead of accepting through it (#305)

[EdmondDantes]

Fixes #305. libuv cannot undo uv_listen, so stop() now pauses: the connection callback leaves the connection unaccepted, and libuv stops watching the socket (Unix) or posts no new AcceptEx (Windows). The first start() runs uv_listen; a later one unpauses and delivers the held connection from a uv_idle_t on the next tick. The delivery stops if a callback pauses or closes the listener. The first stop() switches a Windows TCP listener to single accept, so a listener that is never stopped keeps libuv's default. A failed uv_accept on Unix re-arms the watcher with uv_listen, since libuv restarts it only after a successful accept.

Reviewed before code by a Critic and a Sage (rejected: a uv_poll accept loop), and on the diff by a Critic; its four findings are addressed in this commit.

Evidence:

• sapi/asynctest/tests/test_listen_pause.c (true-async/php-src PR on branch listen-pause-test): 5 of 5 cases red on main, green here, 3 of 3 runs.
• true-async-server h1/062, h1/063, h1/064: 503 on main, 10 of 10 green here; the server's full phpt suite, 516 of 516 green (25 skipped).
• php-async phpt: 1214 of 1220; the 6 failures fail identically on main on this machine. They include php-src helpers through ../../../../, which misses when ext/async is a symlink.

Not measured: the accept rate of a Windows listener after its first pause.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[EdmondDantes]

ASan run of sapi/asynctest/tests/test_listen_pause.c against this branch (minimal ZTS debug build with --enable-address-sanitizer, ASAN_OPTIONS=detect_leaks=1 USE_ZEND_ALLOC=0): 5 of 5 cases, 3 of 3 runs, no AddressSanitizer or LeakSanitizer report; test-listen_uaf and test-timer_overflow clean as well. Known-answer check: with the pefree in listen_close_cb removed, LeakSanitizer reports the 528-byte listen event from listen_on_free_port.