PL-135522: log HTTPException errors at ERROR level before returning response

[elikoga]

PL-135522

Removing a model from inference but not the gateway makes the system 500 forever.

The update_manifest handler raises HTTPException(500) when the gateway requests models the inference server doesn't know about. FastAPI's built-in handler catches this before the custom @app.exception_handler(Exception) gets it, so the error detail was silently swallowed -- only visible in the bare status code in access.log.

Fix:

• Add log.error() in manager.py before raising the HTTPException, including the unknown, requested, and configured model lists.
• Add @app.exception_handler(HTTPException) in the inference app that logs any HTTPException with status >= 500 at ERROR level with the full detail, then returns the standard error response.