Missing finish_reason check in pd_disaggregation for trtllm

### 🚀 Feature Description and Motivation

TRT-LLM's proxy does `_verify_ctx_response()`(https://github.com/NVIDIA/TensorRT-LLM/blob/ae76df1615e0710c6b97a62b2e31c36c3e0f550c/tensorrt_llm/serve/openai_disagg_service.py#L339) which checks: if finish_reason is NOT "length" or "not_finished", skip the generation phase entirely (context already completed within max_tokens=1). The current code always proceeds to decode regardless.

### Use Case

We do have some tasks only generate 1 token, for example, yes or no, numbers etc. In this case, it may throw errors. Let's make sure the workflow cover such cases.

### Proposed Solution

_No response_