Tensorneko

Features:
- `toml` read and write in `tensorneko.io`.

Fix:
- CPU memory leakage in training

Features:
- Adapt `lightning` to `2.0.*`
- Support `PyTorch` `2.0.*`
- Support Python 3.11

Fix:
- Improve the traceback information for `reduce_dict_by` and `summarize_dict_by`
- Fix the docs mismatch for FaceDetectors

Features:
- `resize_image` and `resize_video` supports `str` type resize method input.
- Set `benchmark` default as `True` for `NekoTrainer`
- Add `indent` and `ensure_ascii` args for `tensorneko.io.write.json`
- Now the items in returned value of `get_parser_default_args` can be modified.

Fix:
- More precise for `tensorneko.io.read.video` with `torchvision` backend.
- `tensorneko.io.write.video` doesn't work with `torchvision` backend.

Features:
- `download_file` for downloading any file from URL with progress bar.
- `SystemStatsLogger` callback to log cpu and memory usage.
- `GpuStatsLogger` callback to log gpu utilization and gpu memory utilization.
- `NekoTrainer` now auto load `SystemStatsLogger` if `gpumonitor` is installed, and `SystemStatsLogger` if `psutil` is installed.
- Rename `util.func` as `util.misc`; `util.dispatched_func` as `util.dispatched_misc`
- `WindowMerger` for aggregating sliding window results.

Features:
- `AnimeFaceDetector` used for detecting faces in anime-style images
- `FID` methods for adding video and image now can be invoked as chains
- The `face_sdk_path` arg for `FaceXZooFaceDetector` can be ommitted after the FaceXZoo repository is auto installed.
- More clear debug information for positional encodings.

Fix:
- `ssim_video` and `psnr_video` cannot use ffmpeg for evaluation
- `ssim_video` and `psnr_video` failed when directly read video from file