feat: implement graceful model discovery for vLLM provider

[akram]

• Attempt model discovery first for backward compatibility
• If discovery fails and refresh_models=false, continue without error
• If discovery fails and refresh_models=true, fail hard with ValueError
• Supports dynamic token authentication scenarios

Fixes authentication issues when vLLM service requires dynamic tokens

What does this PR do?

Implement graceful model discovery for vLLM provider.

Test Plan

LLAMA_STACK_LOGGING="all=debug" VLLM_URL=https://my-vllm-server:8443/v1  MILVUS_DB_PATH=./milvus.db INFERENCE_MODEL=vllm uv run --with llama-stack llama stack build --distro starter --image-type venv --run

and use the following configuration:

providers:
  inference:
  - provider_id: ${env.VLLM_URL:+vllm}
    provider_type: remote::vllm
    config:
      url: ${env.VLLM_URL:=}
      max_tokens: ${env.VLLM_MAX_TOKENS:=4096}
      api_token: ${env.VLLM_API_TOKEN:=fake}
      tls_verify: ${env.VLLM_TLS_VERIFY:=true}
      refresh_models: ${env.VLLM_REFRESH_MODELS:=false}

models:
- metadata:
    display_name: vllm
  model_id: vllm
  provider_id: vllm
  model_type: llm
      

start the server, you should see:

WARNING  2025-10-03 20:40:41,114 llama_stack.providers.remote.inference.vllm.vllm:443 inference::vllm: Model verification failed for model vllm/vllm
         with error <some error>
WARNING  2025-10-03 20:40:41,115 llama_stack.providers.remote.inference.vllm.vllm:444 inference::vllm: Continuing without live check
         (refresh_models=false).

[leseb]

Closing for #3677