[Feature] Sweep P/D deployment and network dimensions in Vidur config optimizer

Motivation

Vidur can model P/D separation through pd_node_ratio and exposes P/D, RDMA and NVLink bandwidth settings, but the config optimizer currently sweeps only model, trace, cluster, scheduler, TP, PP and batch size. As a result, users must hand-run every P:D resource split and network point, and the existing Pareto analyzer cannot directly produce a P/D network phase diagram.

Proposed scope

Add an optional pd_networks dimension to config-explorer YAML. Each point contains a stable name, P-node ratio, P/D bandwidth, RDMA bandwidth, NVLink bandwidth and transfer dtype. The generated values should flow into existing ReplicaConfig arguments and become part of the job/cache key.

Requirements:

• legacy YAML without pd_networks keeps one mixed-serving job and its existing key;
• invalid ratios, zero/negative bandwidth and splits producing zero P or D replicas are rejected before simulation;
• each network point produces a distinct reproducible job key;
• existing capacity search and Pareto analysis consume the generated runs unchanged.

This is a design-space generation feature, not a claim that a single bytes/bandwidth model is sufficient for network tail latency.