Update vllm_model pyproject.toml

[hwolff99]

Fixes #549

[hwolff99]

@pjin-nvidia noted:

we do need vllm as a dependency in order to serve models using gym itself for RL
however it is not the intention for vllm to add all the extra nvidia-cuda-* python package dependencies on a CPU-only target, e.g. macos

[hwolff99]

Closing in favor of #552