vllm test cleanup

[planetf1]

Continues work from #397 and #326.

Problems:

• vLLM tests fail when GPU unavailable instead of skipping gracefully
• Multiple vLLM tests cause CUDA out-of-memory errors when run sequentially
• Tests fail cryptically when Ollama not running instead of skipping
• No CLI options to selectively skip backend tests during development
• pytest hooks throw deprecation warnings and duplicate option registration errors
• Examples in docs/examples/ fail when required backends unavailable
• Token limit errors in vLLM structured output tests cause intermittent failures