Also use a lock file per host, rather than one global file lock.
Commit 9c0275a879 introduced a bug where the lock file was only being
unlocked by the child PID of the resulting fork done in ansible-connection.
This causes delays when a large inventory causes a lot of contention on
that global lock. This patch fixes the problem by ensuring the lock is
released regardless of the fork condition, and also to use a lock file
based on the remote address of the target host, removing the global lock
bottleneck.
Fixes#38892
(cherry picked from commit 7ce9968ce1)
4d0422d4ed