[Async] Fix macOS CTS test flakes: dangling stack ops, RST detection, kqueue event loss (#23570)

Turns out, macOS has some... quirks.

Three categories of fixes for macOS socket CTS test failures:

proactor.c: Track `any_operation_removed` through
process_operation_chain to skip event_set_modify when the fd chain is
unchanged after multishot MORE continuations. On macOS kqueue, redundant
EV_ADD resets the knote's pending state, causing level-triggered events
to be missed until a new state transition occurs.

multishot_test.cc: Fix dangling stack-local operation pointers in all
four multishot tests. Operations submitted inside loops (connect_op,
send_op) now wait for their specific completion tracker instead of
relying on PollUntil returning on any completion. All tests also gain a
safety DrainPending() when the cancel-wait timeout expires without the
multishot terminating, preventing TearDown from processing operations
against destroyed stack memory.

error_propagation_test.cc + socket_test_base.h: Add ReleaseWithRst()
helper that closes a socket with SO_LINGER(0,0) (forcing RST instead of
FIN) and submits a recv probe on the peer to detect the RST and set
sticky failure deterministically. Four error propagation tests converted
from non-deterministic send-until-fail loops to single deterministic
assertions. MultishotRecvPeerCloseDeliversFinalCallback fixed with the
same release+poll-until-final+safety-cancel pattern as multishot tests.

Co-authored-by: Claude <noreply@anthropic.com>