feat: support prefix caching and chunked prefill for deepseek v32 on mlu.

[a120092009]

No description provided.

[gemini-code-assist]

Code Review

This pull request introduces support for prefix caching and chunked prefill for the deepseek_v32 model on MLU hardware. The changes are well-implemented and include several key improvements. A new reshape_from_cache kernel is added to support gathering KV cache data for chunked prefill. The IndexerImpl is refactored for better clarity and to accommodate the new chunked prefill logic. The data parallelism handling in LLMEngine is made more robust to correctly manage mixed forward types across different ranks. Additionally, safeguards are added to prevent enabling these new features on unsupported model variants, and comprehensive unit tests are included to validate the new functionality. The code quality is high, and the changes appear correct and well-tested.