feat: TCP keepalive, reverse-channel probe, list_models cache, .pth atomic write, launch exemption

[m199369309]

Summary

• Require xoscar>=0.9.6 for native TCP keepalive support (BUG: enable TCP keepalive for SocketClient to detect dead connections xoscar#189), removing the xinference-side monkey-patch
• Add supervisor→worker reverse-channel ping probe in _check_dead_nodes to detect dead reverse channels even when worker→supervisor heartbeat is healthy
• Add list_models cache degradation: return cached results when a worker is temporarily unreachable instead of returning empty dict (prevents UI from showing no models)
• Reduce XINFERENCE_GET_MODEL_RPC_TIMEOUT from 120s to 30s to limit actor lock contention during worker communication failures
• Fix .pth file concurrent write race in virtualenv setup by using atomic write (os.replace) and skipping writes when content is unchanged
• Exempt workers from reverse-channel dead detection while launching models, as long-running model downloads can starve the actor event loop causing false-positive dead detection
• Limit huggingface_hub download concurrency to 2 (XINFERENCE_MODEL_DOWNLOAD_WORKERS) to reduce GIL contention during model downloads

Changed Files

File	Change
setup.cfg	xoscar>=0.9.3 → xoscar>=0.9.6 (native TCP keepalive)
xinference/core/supervisor.py	Reverse-channel probe, list_models cache, _workers_launching tracking
xinference/core/worker.py	ping() method, .pth atomic write, download concurrency limit
xinference/constants.py	RPC timeout 120→30, new XINFERENCE_MODEL_DOWNLOAD_WORKERS

Test plan

• CI pipeline passes
• Verify TCP keepalive active via xoscar native: ss -tnop | grep <supervisor-port> shows timer:(keepalive,...)
• Verify reverse-channel detection: simulate network partition triggers "Worker reverse-channel timeout" within 25s
• Verify list_models degradation: worker unreachable → UI still shows cached models
• Verify launching exemption: deploying large model does not trigger false "Worker dead"
• Verify .pth race fix: multi-replica deploy on existing venv succeeds on first attempt

[qinxuye]

LGTM