vllm: Temperature validation silently passes NaN and +Inf values

Temperature validation in sampling_params.py uses comparison operators that silently pass NaN and +Inf values, leading to undefined behavior or CUDA errors in GPU sampling kernels.

All vllm users who allow user-supplied temperature values without proper validation.

Apply the fix from PR #45116 to add math.isfinite() check in _verify_args() and reject non-finite float values.