feat(vllm): support mixed DP launch modes

[esmeetu]

Summary

• Add role-specific prefill_dp_launch_mode, decode_dp_launch_mode, and aggregated_dp_launch_mode overrides for the vLLM backend.
• Allow per-node and per-GPU DP process layouts to coexist in one deployment.
• Share topology port allocation across both layouts so KV-event, NIXL, DP-RPC, HTTP, and system ports remain non-overlapping.
• Derive Dynamo health expectations from the generated process topology.

Dependency

This is a stacked follow-up to #90. PR #90 now contains only the global opt-in dp_launch_mode: per_node implementation; this PR's intended review diff is the top commit, feat(vllm): support per-role DP launch modes.

Until #90 merges, GitHub will also show its prerequisite commit in this PR. After #90 lands, this branch will be rebased onto main so only the mixed-DP changes remain.

Motivation

Some vLLM roles need one process per node so node-local DP ranks share a CUDA namespace, while another role may need independent per-GPU Dynamo workers. The override is process-topology flexibility; it does not claim either layout has an inherent request-ingest performance advantage.

backend:
  type: vllm
  dp_launch_mode: per_node
  decode_dp_launch_mode: per_gpu
  vllm_config:
    prefill:
      data-parallel-size: 8
      data-parallel-hybrid-lb: true
    decode:
      data-parallel-size: 16
      data-parallel-hybrid-lb: true

Unset role overrides inherit the global dp_launch_mode.

Validation

• make check
• Ruff and formatting pass.
• Full test suite: 867 passed, 2 skipped, 6 deselected.
• Mixed prefill-per-node/decode-per-GPU launch has been exercised on a 3xPDEP8 + 1xDDEP16 DeepSeek V4 Pro deployment.