Use the nvidia-supplied nvidia-ml-py instead of pynvml

[mattip]

https://github.com/NVIDIA-Merlin/core/blob/e1eaf2694b402bf2ea99a78d3953d59e8a00e35c/requirements.txt#L16

This project depends on pynvml, where the nvidia-blessed bindings to nvml are in the nvidia-ml-py package instead. This is causing problems for the gpustat project, which needs some internals from nvidia-ml-py to properly support older nvml version. Would it be possible to switch?