[3.13] gh-115634: Fix ProcessPoolExecutor deadlock with max_tasks_per_child (GH-14090...

[3.13] gh-115634: Fix ProcessPoolExecutor deadlock with max_tasks_per_child (GH-140900) (#152928)

The idle worker semaphore counts task completions, not idle workers, so
it can hold a stale token released by a worker that later exited upon
reaching its max_tasks_per_child limit. The worker replacement path
consumed such tokens and skipped spawning a replacement, deadlocking
the remaining queued tasks once no workers were left.

Replace dead workers based on len(self._processes) without consulting
the semaphore. The submit() path is unchanged, preserving on-demand
spawning and idle worker reuse.

Replace the documentation note added in GH-140897 with a versionchanged
entry now that the bug is fixed.

Based on a fix proposed by Tabrez Mohammed.

(cherry picked from commit b706767d8fd7d21afc3f156fb9c173bc99855e0e)