feat(vllm): support direct aggregate serving

[kyleliang-nv]

Summary

• add frontend.type: vllm for aggregate jobs that launch vllm serve directly, without Dynamo
• validate the supported topology: vLLM backend, aggregate-only, one worker on one node, and one frontend
• wire direct-vLLM health checks, telemetry, AIPerf server metrics, profiling, and public-port ownership
• preserve current main per-node DP behavior for Dynamo while keeping direct vLLM in the standard one-process topology

Motivation

Aggregate vLLM benchmarks do not need Dynamo as an orchestration layer. This provides a native direct-serving path while leaving existing Dynamo aggregate and disaggregated configurations unchanged.

Validation

• uv run ruff check on all changed files
• uv run pytest -q tests/test_configs.py tests/test_frontends.py (162 passed)
• uv run pytest -q (893 passed, 2 skipped, 6 deselected)