Replace threading.Lock with asyncio.Lock in forecaster

[philyuchkoff]

Problem

Torch is re-entrant but produces nondeterministic garbage under memory pressure when called concurrently from multiple threads. The previous threading.Lock serialised inference correctly, but at the cost of blocking the entire thread — health checks, queue handling, and other requests all piled up behind a single forecast.

Solution

• TimesFMModel.forecast() is now async def forecast() using asyncio.Lock instead of threading.Lock
• server.py: endpoint is now async def and uses await model.forecast()
• Same correctness guarantees: serialisation at the model level preserves safe Torch behaviour
• rest of the server (healthz, readyz, /v1/model, the async lifespan) is never behind the lock

What was not done

• Not a worker pool — this project benefits from preserving a single model, and several workers would mean multiple model instances and complexity not needed
• No changes to the Go controller (the forecaster as a separate process keeps the concern separate)

Verification

All 31 Python tests pass (uv run --extra dev pytest -q).