fix(evaluator): stop timed-out evaluations from hanging the worker

[Ghosts6]

Fixes a hang where, once a program evaluation times out, the worker process never comes back matching #399 ("Hang after evaluation timeout").

Root Cause

Each cascade/direct evaluation stage runs via asyncio.wait_for(loop.run_in_executor(None, ...), timeout=...). On timeout, wait_for only stops waiting the blocking evaluate/evaluate_stageN call keeps executing in the loop's default executor thread, since Python cannot force-kill a running thread.

process_parallel.py calls asyncio.run(...) once per iteration, and asyncio.run()'s cleanup blocks inside shutdown_default_executor() until every thread in that default executor finishes. So a single timed-out evaluation (e.g. an evolved program with an infinite loop) hangs the entire worker process, even though the timeout itself already returned a result.

Fix

openevolve/evaluator.py:

• Evaluator now creates its own dedicated ThreadPoolExecutor and passes it explicitly to all four run_in_executor calls (direct evaluate + stage1/2/3), instead of None.
• Since it's never the loop's default executor, asyncio.run() no longer waits on it, so a timed-out call's orphaned thread can't block cleanup.

Testing

• python -m unittest discover tests — 431 passed (was 430; total runtime also dropped 65.5s → 31.5s, since other timeout tests were silently blocked on the same issue).
• Added test_timeout_does_not_hang_process_on_cleanup in tests/test_evaluator_timeout.py, which measures wall-clock time around asyncio.run() itself (mirroring how process_parallel.py invokes the evaluator). Confirmed it fails (~8s) on the unpatched code and passes (~3s) with the fix.
• black --check clean on both changed files.

[CLAassistant]

All committers have signed the CLA.