fix(vllm): filter None kwargs before passing to SamplingParams

[andreafresa]

Summary

Fixes #21371

llama-index-llms-vllm hardcodes best_of in _model_kwargs, but vLLM >= 0.19.0 removed best_of from SamplingParams. Since _model_kwargs unconditionally included it, every .complete() call raised:

TypeError: Unexpected keyword argument 'best_of'

Root cause — base.py line 257:

return {**base_kwargs}  # passes all keys including best_of=None

Fix — filter out None-valued keys before returning:

return {k: v for k, v in base_kwargs.items() if v is not None}

This way any parameter whose default is None (including best_of) is simply omitted, making the integration robust to future vLLM API removals without breaking explicitly set values.

Test plan

• Vllm.complete() no longer raises TypeError with vLLM >= 0.19.0
• Parameters explicitly set by the user (e.g. temperature=0.7) are still forwarded correctly
• best_of is still forwarded when explicitly set to a non-None value (for older vLLM versions)