test(e2e): add SIGTERM-mid-run abort coverage to test-api-e2e-cleanup-check.sh

[Dumbris]

Context

PR #1368 (`fix: address final review of test-api-e2e.sh cleanup scoping`)
hardened `scripts/test-api-e2e.sh`'s `cleanup()` trap to reap only this
run's own processes, and its regression check
(`scripts/test-api-e2e-cleanup-check.sh`) now:

• starts a real decoy `mcpproxy serve` core AND a launcher-fixture
  argv stand-in, and asserts both survive a full run of the suite;
• asserts both decoys match the OLD blanket `pkill -f` patterns this
  PR removed, so the survival assertions actually prove something;
• continuously samples (once per second) the full descendant tree of
  the suite's own core (and its audit-log sub-instance, if that
  sub-test runs) while the suite is in progress, and asserts none of
  those descendants outlive the run.

Gap

The check only exercises the suite's normal, complete-to-the-end
run. It does not cover the scenario where cleanup hygiene matters
most: the suite's own process is SIGTERMed mid-run, e.g. partway
through `test_launcher_lifecycle`, with the core, the launcher-test
fixture and the "everything" MCP server's npx child all still live.

A prior draft of the check attempted a two-mode ("complete" + "abort")
design, timed to interrupt the launcher-lifecycle test specifically.
It was simplified away during review because a correctness fix
shouldn't ship blocked on a flaky, timing-dependent second test mode
(see PR #1368 review discussion). This issue tracks adding it back
properly, as separate, focused follow-up work:

• run the suite in the background, using a reliable marker (e.g. a
  fixed sleep after a specific log line, or a semaphore file the
  suite touches at a known point) rather than a raw wall-clock delay,
  to fire SIGTERM once the launcher-lifecycle test has started;
• reuse the existing continuous descendant sampler (already added by
  PR fix(e2e): scope test-api-e2e.sh cleanup to this run's own processes #1368) to assert zero leaks after the abort, exactly as the
  complete-run path does today;
• iterate on it locally until it is not flaky (run it back-to-back
  several times) before merging, since this script is not part of CI
  (it needs a real, slow, external-dependency-having E2E run) and a
  flaky abort mode would erode trust in the whole check.

Not blocking any release; this is a coverage improvement to a local,
manually-run regression check.