Py3nvml

Minor update.

- Fixed some alignment issues with long PIDs in py3smi.
- Added ability to call `py3nvml.grab_gpus` with `num_gpus=-1`. This will grab all available GPUs. Could previously do this by setting `num_gpus` to a large number but this would throw a warning if it couldn't get all available gpus.

To try and keep py3nvml somewhat up-to-date with the constantly evolving nvidia drivers, I have done some work to the py3nvml.py3nvml module. In particular, I have updated all the constants that were missing in py3nvml and existing in the NVIDIA source as of version 418.43. In addition, I have wrapped all of these constants in Enums so it is easier to see what constants go together. Finally, for all the functions in py3nvml.py3nvml I have copied in the C docstring. While this will result in some strange looking docstrings which will be slightly incorrect, they should give good guidance on the scope of the function, something which was ill-defined before.

Finally, I will remove the py3nvml.nvidia_smi module in a future version, as I believe it was only ever meant as an example of how to use the nvml functions to query the gpus, and is now quite out of date. To get the same functionality, you can call nvidia-smi -q -x from python with subprocess.

Updated script to not use a big xml query but rather multiple small queries. Can now handle if a gpu falls off the bus too, still displaying info for the remaining gpus.

Added py3smi script