fix: round-robin DP rank assignment for service discovery mode

[raviguptaamd]

Summary

When using ZMQ service discovery with --intra-node-data-parallel-size > 1, worker addresses arrive without @rank suffixes. The router left prefill_dp_rank as None, so:

• The prefill request lacked X-data-parallel-rank header
• The decode kv_transfer_params omitted remote_dp_rank

This caused the MoRI-IO WRITE mode handshake to always target DP rank 0, deadlocking all other DP ranks.

Fix

Add an atomic round-robin counter (discovery_dp_rank_counter) to VllmPDRouter that assigns DP ranks when extract_base_http_and_dp_rank returns None and DP size > 1. The assigned rank flows into:

1. Prefill request's X-data-parallel-rank header
2. Decode request's remote_dp_rank in kv_transfer_params

This matches the behaviour of the toy proxy which already does round-robin DP rank assignment.

Changes

• src/routers/http/vllm_pd_router.rs:
  • Add discovery_dp_rank_counter: AtomicUsize field to VllmPDRouter
  • Initialize to 0 in both constructor paths (discovery and non-discovery)
  • In process_vllm_two_stage_request_discovered(), generate round-robin DP rank when prefill_dp_rank.is_none() and intra_node_data_parallel_size > 1

Test plan

• 1P/1D DeepSeek-V3-5layer on MI300X (8 DP ranks per node, MoRI EP, WRITE mode) — warmup and benchmark pass with correct per-rank MoRI-IO handshakes
• Unit tests for round-robin assignment with DP size > 1
• Verify no regression for non-discovery (static URL) mode where @rank suffixes are present

Builds on top of #157 (MoRI WRITE mode concurrent dispatch).