When stopping an instance in the cluster tests, disable appendonly first, so that SIGTERM won't be ignored.

Recently in #9679 i change the test infra to use SIGSEGV to kill servers that refuse
the SIGTERM rather than do SIGKILL directly.

This surfaced an issue that i've added in #7725 which changed SIGKILL to SIGTERM (to resolve valgrind issues).
So the current situation in the past months was that sometimes servers refused the
SIGTERM and waited 10 seconds for the SIGKILL, and this commit resolves that (faster termination).

Fix failures introduced by #9695 which was an attempt to solve failures introduced by #9679.
And alternative to #9703 (i didn't like the extra argument to kill_instance).

Reverting #9695.
Instead of stopping AOF on all terminations, stop it only on the two which need it.
Do it as part of the test rather than the infra (it was add that kill_instance used `R`
to communicate to the instance)

Note that the original purpose of these tests was to trigger a crash, but that upsets
valgrind so in redis 6.2 i changed it to use SIGTERM, so i now rename the tests
(remove "kill" and "crash").

Also add some colors to failures, and the word "FAILED" so that it's searchable.

And solve a semi-related race condition in 14-consistency-check.tcl