[codex] Block legacy fine-tuned SAM3 loads

[hansent]

Summary

• block fine-tuned SAM3 legacy loads during Roboflow model type resolution when SAM3_FINE_TUNED_MODELS_ENABLED is disabled
• return a dedicated 501 deployment-level error instead of falling through to a generic failure after model construction
• add regression coverage for both the registry guard and the HTTP error mapping

Why

The legacy /{dataset}/{version} inference path could resolve a fine-tuned SAM3 model, instantiate SegmentAnything3, and download large SAM3 weights even on deployments where fine-tuned SAM3 loading is disabled. Inference then failed later because that endpoint cannot actually serve SAM3. This change short-circuits that path before model construction and weight download.

Impact

Serverless or otherwise restricted deployments now reject fine-tuned SAM3 model IDs earlier and more clearly, while core sam3/... models continue to work as before.

Validation

• PYTHONPATH=/Users/hansent/.codex/worktrees/e7e5/inference/inference_models:/Users/hansent/.codex/worktrees/e7e5/inference uv run --no-project --with-requirements requirements/requirements.test.unit.txt pytest tests/inference/unit_tests/core/registries/test_roboflow.py tests/inference/unit_tests/core/interfaces/http/test_error_handlers.py
• result: 65 passed

[PawelPeczek-Roboflow]

As stated in comment - if you have time, you can extend to other models, but since this is a good start, we can merge in this form

[PawelPeczek-Roboflow]

I believe this could work in much broader context - i.e for all models
but that's a good start so I am willing to approve