[artem-sitd]

Fix #386.

Problem

With --workers > 1, SIGTERM/SIGINT torn down in-flight requests:

• idle workers exit immediately once shutdown_event is set;
• the master is blocked in wait(process.sentinel ...), which returns as soon as any worker exits;
• the loop then ends (because active is False) and the master calls process.terminate() on the remaining, busy workers.

The result is that requests still being served are cut off ("Empty reply from server"), even when -w 1 serves them fine.

Change

Replace the terminate() loop with _drain(processes), which joins the remaining workers so they can finish their own graceful shutdown (bounded on the worker side by config.graceful_timeout). This matches the single-worker behaviour and mirrors the existing unbounded process.join() in the reload handler.

Tests

Added tests/test_run.py asserting the master joins workers and never terminates them. Existing test_utils.py and test_config.py pass.

Note: on Python 3.12+, worker-side draining can be effectively unbounded because of Server.wait_closed() (#308); that is a separate issue and its fix would let config.graceful_timeout (or a future hard cap in the master) be applied without cutting in-flight requests.