[Feat] Support TensorRT-LLM generation-first (parallel) P/D scheduling

### 🚀 Feature Description and Motivation

low priority since context-first mode has been supported

The pd router currently only supports context-first scheduling (sequential: prefill → wait for response → decode).

TRT-LLM also supports generation-first mode where the proxy sends context_only and generation_only requests in parallel — the GEN worker enters `DISAGG_GENERATION_INIT` state, uses ctx_info_endpoint + disagg_request_id to pull KV cache directly from the CTX worker while it's still computing, reducing end-to-end latency.

Implementation requires: caching `/server_info` from CTX workers at router init (returns ctx_info_endpoint and ctx_dp_rank), dispatching both requests concurrently via goroutines, injecting the cached server info into the GEN request's disaggregated_params, and setting schedule_style:

> Note that ctx_dp_rank is the attention data-parallel rank — relevant when CTX workers use data parallelism (TP x DP). For single-instance deployments it's 0, but for multi-instance setups the router needs to track which DP rank each CTX worker exposes and match correctly when routing.

### Use Case

Support generation-first strategy

### Proposed Solution

_No response_

Start by locating the pd router's context-first request flow and its /server_info handling. Trace how CTX and GEN requests are routed, then define completion as concurrent dispatch, cached ctx_info_endpoint and ctx_dp_rank propagation, correct disaggregated_params, and generation-first scheduling with multi-DP routing support.

Written by the indexing model from the issue text.