fix: preserve generated vision token text embeddings

[dingzhiqiang]

When an autoregressive multimodal sequence contains a generated image/video special token, matching visual embedding positions by token ID alone also selects that generated token. This can misalign visual features or fail during embedding scatter.

Use optional mm_token_type_ids (0=text, 1=image, 2=video) to select real visual input positions while preserving generated special-token IDs, text embeddings, and gradients. Validate shape, modality, placeholder IDs, payload presence, and feature counts. Callers without modality types retain token-ID-based selection, with explicit feature-count validation.

Validation:

• CPU regression coverage for image/video embedding placement and gradients, legacy callers, invalid metadata/payloads, and image-only configurations.
• 55 tests passed in the combined bridge, AReaL mRoPE, and packed-input regression suite.
• Full bridge pre-commit checks passed.
• A downstream multimodal RL run completed its next optimizer step and saved an optimizer-inclusive recovery checkpoint with this change.

Scope: callers must supply correct input-aligned modality types to distinguish generated special tokens. This change only covers the shared HuggingFaceVit embedding helper; it does not modify generation or model-specific mRoPE handling.