reset cache for mm models

[hjh0119]

Add two vllm reset cache interfaces for multi-modal models, and update cookbooks

[gemini-code-assist]

Code Review

This pull request introduces methods to reset multimodal and vision encoder caches in the vLLM sampler, ensuring memory is freed after profiling and stale embeddings are cleared after weight updates. The changes are integrated into the GKD and GRPO cookbooks. Feedback includes a suggestion to add a safety check for the reset_mm_cache method to prevent errors on older vLLM versions, a correction for a version number typo in the docstrings, and a recommendation to add the @remote_function decorator to the existing reset_prefix_cache method for consistency in distributed environments.