Tensorneko

Fix:
- Training error when using `lightning` 2.1
- Docs are not clear for `VideoReader`, `RoundRobinDataset`, `NestedDataset`, `ListDataset`.

Features:
- Better type hint for eventbus and dispatch
- Dynamic import for `mita_tqdm`
- `on_step` and `on_epoch` arguments for `GpuStatsLogger` and `SystemStatsLogger`
- `Timer` now doesn't require parentheses as decorator

Fix:
- `crop_with_padding` using numpy array
- `sample_indexes` not imported
- `FID` not working with `progress_bar=True`
- import error with pytorch lightning 2.1.3

Features:
- Add support for pytorch 2.1 and lightning 2.1
- Auto infer for `notebook.display`, now supports audio, video, image, and code.
- `ListDataset`, like `TensorDataset` in `torch` but directly uses list
- `SequentialIterSampler`, can split dataset to multiple epochs. Normally can be used for training in large-scale dataset but want to utilize the learning rate scheduler based on epochs.
- `BinaryClassifier`. A predefined model for binary classification.
- `sample_indexes`. Sample a continious indexes giving a source sequence length.

Fix:
- Import problem raise by `RoundRobinDataset`
- `FID` calculation problem when set `num_workers > 0`.
- `iou_2d` calculation error

Features:
- `Timer` now can disable the auto print, for better manual control.
- `Read` and `Write` auto infer format
- `try_until_success` for trying a function with multiple times if it raises exceptions.
- `DummyLogger` for disabling annoy logging information when debug

Fix:
- The unusual argument order in `MatWriter`

Features:
- Update `eventbus` api for less import.
- `load_py` function in `util.misc` for easiler import a py file as a module.

Fix:
- Train loss is not displayed in the progress bar
- Default audio backend doesn't work for audio writer and reader
- FFmpeg -ss and -i wrong order for clipping videos
- Add warn for if the input of video writer is from 0~255 unit8

Features:
- Add ffmpeg backend for audio reader

Fix:
- Tensorboard graph logging not working
- Training will fail in DDP due to a bug in gpu logging callback