fix(v1): guard the pool health reply so a departed client can't crash the pool

[dumko2001]

TL;DR — under ROUTER_MANDATORY=1, one client abandoning a health poll crashes the entire env-server pool (all workers, all in-flight rollouts). Three-line fix.

Repro

Client calls .health() on a short timeout, times out, and closes its DEALER socket in the gap between the broker reading the request (pool.py:196) and sending the reply (pool.py:199). The ROUTER has already dropped that identity, so the send raises zmq.ZMQError.

Impact

run()'s only handler is except (asyncio.CancelledError, KeyboardInterrupt) (pool.py:239), so the ZMQError escapes the loop. It unwinds into finally: self._shutdown() (pool.py:242), which calls w["process"].terminate() on every worker (pool.py:249). The whole pool goes down, killing every concurrent rollout, not just the health request that triggered it.

Fix

# before
if method == b"health":
    await self.frontend.send_multipart([client_id, request_id, _HEALTH])

# after
if method == b"health":
    with contextlib.suppress(zmq.ZMQError):
        await self.frontend.send_multipart([client_id, request_id, _HEALTH])

The worker-reply send two branches down already uses this exact guard (pool.py:235); the health send was the one path missing it. Live clients are unaffected, since the suppress only swallows a send to a peer that's already gone. Happy to add a regression test that drops a client mid-health poll.

Reopens #2071.