Improve out-of-the-box recipe for Kimi-K2.5

[faradawn]

Recipe file: moonshotai/Kimi-K2.5.md

Goal: 1) no arbitrary flags and environment variables, 2) vllm serve just works -the flags should be pushed to the core engine codebase.

Action items

• --compilation_config.pass_config.fuse_allreduce_rms true: Can we make default for MoE models on Hopper hardware? -> Yes. That is for v0.17. The latest vllm does not need these flags. (Hanjie Qiu, Wei Zhao).
• --mm-encoder-tp-mode data: Can we push this to vllm codebase to auto-detect?
• --tool-call-parser / --reasoning-parser / --enable-auto-tool-choice: Can we parser names from model? -> Not for now. (Ben Browning). But can add this in the future.
• --trust-remote-code: Can we eliminate this?
• Split into Min Latency and Max Throughput configs. -> Found. (Ankur Singh)
• Add validated results (throughput, TTFT, TPOT, ITL) etc. -> Added. (Ankur Singh)
• Fix Eagle3 command to match base serve flags consistently.