feat: support Kimi-K3 DSPARK PD decode radix cache

[qybnb]

Summary

• allow the narrow Kimi-K3 + DSPARK + PD decode-radix combination while keeping the prompt-end KDA/conv state from Prefill authoritative
• reuse Decode-local target/draft token KV prefixes and transfer only the missing KV suffix from Prefill
• keep the Ascend/NPU graph path from selecting the CUDA-only Kimi-K3 verification backend
• add opt-in runtime diagnostics and correct Decode radix cache-pool accounting

State and transfer contract

For prompt length N and Decode-local radix hit L:

• Decode radix owns the per-token target MLA/draft KV prefix [0, L)
• Prefill computes or restores state through N
• Prefill transfers target/draft KV only for [L, N)
• Prefill independently transfers the complete prompt-end KDA/conv state S_N
• Decode does not index KDA checkpoints in its radix tree

The path is intentionally limited to Kimi-K3 + DSPARK with DCP=1 and device-resident cache. HiCache, other speculative algorithms, other hybrid-SSM models, and draft prefetch are outside this change.

Runtime diagnostics

Set SGLANG_PD_RADIX_DIAGNOSTIC_LOG=1 on Prefill and Decode to emit rank-0 diagnostics for Decode prefix matching, Prefill KV suffix transfer, prompt-end KDA/conv state payloads, and final Decode cache accounting. The switch is disabled by default.

---

CI States

Latest PR Test (Base): ⚠️ Run #33570754761
Latest PR Test (Extra): ⚠️ Run #33570754543
Latest PR Test (AMD ROCm 7.2): ⚠️ Run #33570754805