Stop the publisher cluster as well as the subscriber (#470)

[ChronicallyJD]

Closes #470.

The defect

pgc_setup installs trap pgc_teardown EXIT. test/logical_subscriber.sh then installed its own:

sub_cleanup() { pgc_pg "pg_ctl -D '$SUB_DIR' -m immediate stop" >/dev/null 2>&1 || true; }
trap sub_cleanup EXIT

A second trap ... EXIT replaces the first. So the subscriber cluster was stopped and the suite's own cluster never was. The suite passed, every run, and left a live postmaster behind.

Measured, not argued

PG17, from a box reaped to zero orphaned postmasters:

before fix:  logical_subscriber  rc=0  before=0 after=1 delta=1  PASSED
after fix:   logical_subscriber  rc=0  before=0 after=0 delta=0  PASSED

Gated on PG18 and PG19, with replication.sh alongside because it is the suite that already chained correctly and must keep doing so:

PG18  logical_subscriber  rc=0  checks run: 20  PASSED
PG18  replication         rc=0  checks run: 41  PASSED
PG19  logical_subscriber  rc=0  checks run: 20  PASSED
PG19  replication         rc=0  checks run: 41  PASSED

Why it is not cosmetic

The port band is finite. A suite that cannot get a port fails after 8 start attempts with could not create any TCP/IP sockets, which is indistinguishable from a real red and is attributed to whichever suite drew the exhausted port, not to the one that leaked. 37 orphans had accumulated on the dev container and turned two majors of a gate red for reasons unrelated to the code under test.

The audit behind it

objstore_module had the identical defect and was fixed in #446. This is the other one.

The audit had to be measured rather than read, and that mattered: harness_selftest installs its own EXIT trap at line 73, before pgc_setup at line 90, so by inspection pgc_setup should clobber it and the squatter cluster should leak. Measured, it does not.

suite	delta	verdict
harness_selftest	0	no leak, despite the ordering
logical_subscriber	1 -> 0	this PR
replication	0	already correct, chains pgc_teardown
objstore_module	1 -> 0	fixed in #446

Left open deliberately

Nothing in the harness detects this class. A suite can leak a cluster and still report PASSED, which is how both instances survived unnoticed. #470 records that a before/after postmaster count around each suite would catch the whole category rather than these two instances; that is a harness change and not this PR.