Free each thread's strtod Bigint pool

[EdmondDantes]

What was broken

Bfree() does not release a Bigint; it pushes it onto freelist[k], which under ZTS is EG(strtod_state).freelist — one pool per thread. The only thing that empties a pool is destroy_freelist(), reachable solely through zend_shutdown_strtod(), which runs once, on the main thread. executor_globals_ctor zeroes the state for every new thread and executor_globals_dtor never touched it.

So every thread that formatted or parsed a double leaked its pooled Bigints when its TSRM storage went. Measured with a ThreadPool worker doing one sprintf('%6.3f', 1.5):

68 bytes in 2 blocks are definitely lost

— a 36-byte Balloc(1) from d2b() and a 32-byte Balloc(0) from rv_alloc(). Four workers, one float each: 272 bytes in 8 blocks, exactly linear. It is small per thread and unbounded per process: a pool that rotates its workers pays it again on every rotation.

The change

destroy_freelist() and free_p5s() take the state to free instead of reaching through EG(), and executor_globals_dtor() frees the departing thread's pool. The state has to travel as an argument: TSRM runs a thread's storage destructor from whichever thread performs the shutdown, so EG() there belongs to somebody else.

zend_shutdown_strtod() keeps its signature and now goes through the same path for the main thread. Running it before the destructor is safe — the teardown NULLs both lists, so the second pass finds nothing.

The two #undefs before the teardown are needed because freelist and p5s are macros over EG(strtod_state) for the rest of the file, and they would otherwise swallow the field names in state->freelist.

Measurements

	before	after
one pool worker formatting one double	68 bytes definitely lost in 2 blocks	0
four workers, one double each	272 in 8	0

Zend/tests: 5398 passed, 129 skipped, 0 failed. ext/standard/tests/math: 164 passed, 0 failed. Build is ZTS, debug.

Found while chasing a leak in another project against this fork; the same code is in php-src master, so this is not fork-specific.

[EdmondDantes]

Closing: this is upstream php-src behaviour, not a fork concern, and this repository is not where we carry such patches.