Add ability to pass options to VLLM

[robinhad]

This pull request enables passing options to VLLM directly. This is useful for the following scenario:

1. gemma-3-12b-it model, which can fit in 2x24G GPU, but has a long-context of 128k tokens. User can then specify max_model_len directly like:
   VLLM_USE_V1=0 VLLM_WORKER_MULTIPROC_METHOD=spawn lmms-eval --tasks mmmu --model vllm --model_args model_version=google/gemma-3-12b-it,tensor_parallel_size=2,gpu_memory_utilization=0.95,max_images=1,max_videos=0,max_audios=0,max_model_len=4096 --batch_size 100 --log_samples --output_path lmms-results

It enables support for all VLLM engine args.