[mm] fix: missing cached item in beam search

[AlpinDale]

No description provided.

[gemini-code-assist]

Code Review

This pull request aims to fix an issue with missing cached items in beam search for multimodal models by removing a preprocessing step. While this is a valid approach to prevent double expansion of placeholders, it introduces a regression where mm_processor_kwargs are always set to None, ignoring any values passed in the prompt. This could break multimodal beam search for models that depend on these arguments. I have provided a code suggestion to correctly extract these kwargs from the prompt.