feat(vllm): support configurable direct serve binary

[jiahanc]

Summary

• Add backend.vllm_serve_binary for direct aggregate frontend.type: vllm jobs.
• Keep vllm as the default, preserving all existing recipes.
• Allow vllm-rs or an absolute executable path when the Rust binary is not on PATH.
• Leave Dynamo, sidecar, and vllm-router launch paths unchanged.
• Document a managed-engine Rust frontend recipe and cover command construction plus schema round-tripping.

Motivation

The direct vLLM backend currently hardcodes vllm serve. Some vLLM images ship the managed-engine Rust OpenAI frontend as vllm-rs, occasionally only at a package-local absolute path. Recipes need a first-class way to select that executable while retaining srtctl's existing model, topology, port, device, profiling, and vLLM CLI argument construction.

Duplicate-work check

No open PR matched searches for vllm-rs, Rust frontend vllm, or configurable serve binary in NVIDIA/srt-slurm.

Tests

• .venv/bin/python -m pytest tests/ -v (1753 passed, 2 skipped, 6 deselected)
• .venv/bin/ruff check src/srtctl/
• .venv/bin/ruff format --check src/srtctl/ tests/test_configs.py
• .venv/bin/ty check src/srtctl/backends/vllm.py
• git diff --check

AI assistance was used to prepare this change. The submitter is responsible for reviewing and understanding every changed line before merge.