[None][fix] Set token IDs on request after router tokenization to avoid re-tokenization

KvCacheAwareRouter now sets prompt_token_ids (ChatCompletionRequest) or
replaces prompt with token IDs (CompletionRequest) after tokenizing,
so the downstream worker server skips redundant tokenization.

Also adds proper ChatCompletionRequest handling via apply_chat_template.

Signed-off-by: Lizhi Zhou <1432185+reasonsolo@users.noreply.github.com>

Author: reasonsolo

scripts/attribution/data/dependency_metadata.yml

@@ -20,6 +20,10 @@ cutlass/v4.3.0:
   copyright: 501a1dd5925b22dbe66ce8f195440dec
   license: 0d49472f2c1269d573d79e5603778a9f
   source: https://github.com/NVIDIA/cutlass/tree/v4.3.0
+cutlass/v4.4.1:
+  copyright: aa8578535b24304362e63f0e170d8e60
+  license: 1078b45037d7c56c5c2af36977a4d08e
+  source: https://github.com/NVIDIA/cutlass/tree/v4.4.1
 deep_ep/5be51b228a7c82dbdb213ea58e77bffd12b38af8:
   license: 452b3ef002dc6ec283fb723f0dd84997
   source: https://github.com/deepseek-ai/DeepEP/tree/5be51b228a7c82dbdb213ea58e77bffd12b38af8