Convert pcp_worker die handler to flag-only.

I have reviewed the patch(029-pcp-worker-die-flag-only) from Emond.
It has also passed all regression tests.

die() (SIGTERM/SIGINT/SIGQUIT handler in the PCP worker child) directly
called ereport() and proceeded to exit logic from signal context. These
operations are not async-signal-safe, and ereport() in particular can
allocate memory or acquire locks that the main loop may already be
holding, risking corruption or a hang if the signal arrives at the
wrong point.

Convert die() to the canonical pgpool flag-only pattern used elsewhere
in this file. die() now only saves errno, records the signal number
into pcp_worker_shutdown_signal, sets pcp_worker_shutdown_request, and
restores errno before returning. The actual handling, logging the
shutdown request and performing the SIGTERM "smart shutdown" or
SIGINT/SIGQUIT immediate exit, is moved into the new
process_pcp_worker_shutdown_request(), which is called from the main
loop at safe points (before and after each PCP packet read).

Note that in the originally provided patch, the signal was tracked with a single flag,
pcp_worker_shutdown_request, holding the signal number directly.
To align with 026-close-idle-connection-flag-only and 027-pcp-exit-handler-flag-only,
which keep the variable indicating that a signal has arrived separate
from the variable storing which signal it was, I split this into two variables here as well.
For reference, I'm also attaching the original patch(0001-fix-pcp-worker-die-flag-only.patch).

Also, regression tests passed after applying this together with the other
similar signal handler fixes, 026-close-idle-connection-flag-only and 027-pcp-exit-handler-flag-only.
However, 028-lifecheck-exit-handler-flag-only is failing due to a timeout
in the 004.watchdog test case, so 028 has not been posted to this ML yet.

Thank you Emond.

Taiki Koshino<koshino(at)sraoss(dot)co(dot)jp>
SRA OSS K.K.
TEL: 03-5979-2701 FAX: 03-5979-2702
URL: https://www.sraoss.co.jp/