Improve vLLM upstream health checks to only pass when models are servable

[smarterclayton]

As documented in #550, the default vLLM configuration could be improved and documented better. A startupProbe on /health is the right default for vLLM given it does not load the server until a very long model load is complete, but tunables may vary.

[kfswain]

I believe #550 actually resolved this:

gateway-api-inference-extension/config/manifests/vllm/gpu-deployment.yaml

Line 137 in c546506

# vLLM does not start the OpenAI server (and hence make /health available)