fix(mlnode): scheduler-heartbeat liveness for vLLM runner + damped proxy health probe

[baychak]

Problem

vLLM /health is the wrong liveness signal in both directions, and mlnode relies on it twice:

1. Real hang, /health 200. A TP worker deadlocks on the shm collective: scheduler frozen, zero throughput — but AsyncLLM.check_health only reads the errored flag. The node silently serves nothing until someone notices (5–22 min in production).
2. Busy engine, /health timeout. A saturated engine misses the 2 s deadline: runner.is_available() reports dead → a healthy node gets restarted (observed 8+/day); the proxy prober flips on a single miss → the compatibility server stops, dropping in-flight relays (23 flaps/2.3 h).

Fix

• runner.py: liveness = vllm:iteration_tokens_total_count heartbeat — advances on every engine step, freezes only on a stall. Per-instance; re-baselines after engine restarts; aggregates multi-registry scrapes; a failed scrape can't pass for idle; ConnectTimeout = saturation (grace path), refused = dead. MLNODE_HANG_GRACE_SEC (default 120, 0 disables).
• proxy.py: prober timeout 2 → 10 s; unhealthy only after 3 consecutive failures, instant recovery.

Validation

21 unit tests. 4×B300 Kimi-K2.6 fleet, several days: false restarts → 0 (incl. a full hour at max_num_seqs saturation), real stalls detected in ~2 min and self-healed via the existing restart path, proxy flaps → 0. Detection layer only — the underlying vLLM stall is addressed in gonka-ai/vllm#58.