Embedding pool memory admission and OOM recovery

Treat native ONNX OOM as an uncatchable SIGKILL and prevent it through fail-closed admission. Always retain one primary worker. Admit extra workers only from fresh container-aware memory data, reserving memory for admitted-but-unhealthy workers and dividing shared budget by the pool ceiling. Clamp each worker's token cap against its share and current trustworthy free memory on every request; unknown constrained headroom uses the minimum cap, never host-memory guesses. Re-clamp direct post-respawn resubmissions while preserving request IDs. Never increase usable memory during clamping or re-probe upward without genuine improvement or beyond a known failed cap. Keep pool behavior tests separate from dedicated per-worker OOM/respawn tests.