[VLM] Support caching in merged multi-modal processor (outside the HF calling loop)

[VLM] Support caching in merged multi-modal processor (outside the HF calling loop) #1701