fix: flush coverage data on SIGTERM/SIGINT

[bartlomieju]

Summary

Fixes coverage data loss when a Deno process with DENO_COVERAGE_DIR is
terminated via SIGTERM or SIGINT (e.g. by Playwright's webServer teardown,
Docker, systemd, Ctrl+C).

Currently, CoverageCollector::stop_collecting() only runs via the normal exit
path in CliMainWorker::run(). If the process receives a termination signal,
the event loop exits immediately and coverage data is never written to disk.

Approach

When coverage collection is enabled, the event loop in cli/worker.rs uses
tokio::select! to race the normal event loop future against a termination
signal stream. On receiving SIGTERM or SIGINT:

1. Break out of the event loop
2. Skip unload/processExit event dispatch (process is being killed)
3. Call stop_collecting() to flush coverage via V8 inspector
4. Re-raise the original signal with the default handler so the process exits
   with the correct signal exit code

Changes

• ext/signals/lib.rs — Added cross-platform helpers to deno_signals:

  • SignalStreamWithKind — async stream that reports which signal was caught
  • termination_signal_stream() — listens for SIGTERM+SIGINT on unix, SIGINT
    on Windows
  • raise_default_signal(signo) — re-raises a signal with the default handler

• cli/worker.rs — Modified CliMainWorker::run() to intercept termination
  signals when coverage is enabled, flush coverage data, then re-raise the signal

• tests/specs/coverage/sigterm_flush/ — Spec tests verifying coverage is
  written on SIGTERM for both main worker and web worker scenarios

Limitations

• Unix only for SIGTERM (Windows has no cross-process SIGTERM equivalent)
• SIGINT works on both platforms
• Does not handle SIGKILL or Windows TerminateProcess (uncatchable by design).
  A follow-up PR will add periodic coverage flushing to handle force-kill
  scenarios.

Reproduction

The original issue was discovered with Playwright's webServer config:
Playwright sends SIGTERM (with gracefulShutdown) or SIGKILL (default) when
tearing down the dev server. With DENO_COVERAGE_DIR set in webServer.env,
coverage was never written.

Test plan

• Spec test: cargo test sigterm_flush — spawns a Deno server with
  DENO_COVERAGE_DIR, sends SIGTERM, verifies coverage files are written
• Web worker variant: same test but with a web worker, verifies both main
  and worker coverage are flushed (2 files)
• Manual: run deno run --allow-net server.ts with DENO_COVERAGE_DIR set,
  send SIGTERM, verify .json files appear in coverage dir

🤖 Generated with Claude Code

[nathanwhit]

LGTM

[bartlomieju]

Needs a bit more work to flush coverage data periodically to guard against SIGKILL (which can't be handled)