Pytorch-gan-metrics

1. use `tqdm.auto` instead of `tqdm`
2. use argument name instead of positional parameter for `calc_fid_stats`
3. update `version`

- Support `torchvision==0.13.1`.
- Improve documents.
- Make unittest self contained.
- Add wheel to release

1. Fix type hint
2. Fix import error of `load_state_dict_from_url`
3. Add `--use_torch` to cli tool `pytorch_gan_metrics/calc_metrics.py`
4. Update reproducing table in readme

1. `utils.py` somehow disappeared on pypi

- Fix CUDA is reinitialized in subprocess when multiprocess is used for distributed training.
- Add argument `device` with default value `torch.device('cuda:0')` to `core.get_inception_feature` which can be assigned by `**kwargs` for all utility functions `utils.*`.