Drain the :consume stderr tail when the child exits

[woohp]

In :consume stderr mode, the tail can miss the child's last writes when the exit status arrives before the stderr readiness event.

finish_exit/2 serves parked stdout/stderr readers from the pipe (the comment there explains why: the exit status can arrive over the UDS while output is still sitting in the pipe), but the :consume drain is only driven by select events, so nothing drains it at exit. If the UDS exit message wins the race, stderr_tail/1 called right after await_exit/1 returns a short or empty tail. That includes run/2 with stderr: :capture, whose comment says the tail is complete once the process has exited. stderr_tail_test.exs works around it with wait_until_drained/2 ("avoids racing the select-driven consumer against the OS process exit notification").

Fix: also call maybe_consume_stderr/1 in finish_exit/2. The child's write ends are closed by then, so the drain reads the remaining bytes and stops at :eof (or :eagain if a grandchild still holds the pipe open). The :consume_stderr_more handler already treats a drain after exit as a no-op