inference: add health-first backend benchmark helper

[sheawinkler]

Summary

• upgrades the local inference benchmark helper to health-first endpoint probes
• supports vLLM, vLLM Metal, MLX/mtplx, SGLang, llama.cpp, LM Studio, TGI, TensorRT-LLM, OpenAI-compatible, and Ollama endpoints
• keeps generation optional with --chat so the script never pulls or launches models by default

Local verification

• bash -n scripts/benchmark_inference_backends.sh
• git diff --check