fix vllm dp & reset_encoder_cache & fix vllm init with zero3

[hjh0119]

1. fix vllm dp with dense model Major speed degradation in GRPO with async external server after vllm update #9221 [BugFix] Support online dense model DP without overhead vllm-project/vllm#30739
2. support reset_encoder_cache to clean vision embedding cache Support clear mm and encoder cache vllm-project/vllm#33452
3. Fix a bug where, under zero3 + colocate, vllm model initialization may fail due to the zero3 context introduced by training (e.g. gemma4).) Gemma4 Support FFT + Zero3 #9255
4. fix timeout issue in async dp introduced by Fix DP coordinator ZMQ port TOCTOU vllm-project/vllm#37452

[gemini-code-assist]

Code Review

This pull request introduces several enhancements to the vLLM engine integration, including new methods for resetting encoder and multimodal (MM) caches and improved device visibility management for data parallel ranks on both CUDA and NPU platforms. Review feedback identifies potential AttributeError risks for the reset_mm_cache call if it is not supported by the underlying vLLM engine. Additionally, there are suggestions to refactor duplicated logic in the cache reset broadcasting methods and a request to verify the '0.16' version check, which may be a typo for '0.6.x'.

[hjh0119]

/gemini review

[gemini-code-assist]

Code Review

This pull request enhances the vLLM engine integration by increasing the DP coordinator timeout to 180 seconds and improving device visibility management for data-parallel rollout workers. It also introduces functionality to reset encoder and multimodal caches, adding corresponding FastAPI endpoints and client-side methods. The review feedback points out that the implementation of these new asynchronous endpoints uses blocking I/O, which could negatively impact the event loop's performance, and recommends offloading these operations to an executor.