Allow custom OCR model registry entries without forking

[davanstrien]

MODEL_REGISTRY now covers the built-in model set, but adding a private/custom OCR script still requires editing run.py.

Add a supported configuration path (for example YAML/TOML or repeatable CLI flags) that can define at least:

• model slug and model id
• uv-script URL
• default hardware flavor and arguments
• optional container image / Python / environment settings
• display parameter count

Validation should fail before launching Jobs, and built-in registry entries should remain the zero-config default.

Split from #1 after the built-in model expansion was completed.