SUPPORTED_MODALITIES is global, not per-model

[t-avil]

What happened?

Summary

entrypoint.py:34 declares one modality set for every model:

SUPPORTED_MODALITIES = frozenset({"text", "image", "audio", "video", "action", "scalar", "tensor"})

used at entrypoint.py:273 to validate input_modalities on intake. It has
no per-model dimension, so audio validates for BAGEL even though BAGEL has
no audio encoder — the request is accepted and only fails (or misbehaves)
downstream, instead of being rejected at intake with a clear message.

Current state

• One frozenset, checked against every model's input_modalities regardless
  of what that model actually implements.
• Each model already knows its own supported modalities implicitly (e.g. via
  which load_* methods and prefill walks it defines), but nothing exposes
  that to the entrypoint.

Proposed fix

Add a per-model modality declaration (e.g. a SUPPORTED_MODALITIES class
attribute or a supports_modality(name) method on Model), and have
entrypoint.py:273 consult the loaded model instead of the global frozenset.

Acceptance criteria

• A request declaring a modality the loaded model doesn't support is
  rejected at intake with a clear 400, not accepted and left to fail later
• BAGEL specifically rejects audio/video input modalities
• Existing models' currently-valid modality combinations are unaffected

How to reproduce

No response

Environment

No response