Tensorneko

Features:
- Implement `get_parser_default_args` for parsing default argument from `argparse` library.
- Move tensorboard server to `tensorneko_util`

Features:
- `merge_video_audio` function based on ffmpeg
- `resample_video_fps` function based on ffmpeg
- refactor `io` module
- implement`io.read.mat`, `io.read.json`, `io.read.pickle`, `io.write.mat`, `io.write.json`, `io.write.pickle`
- `io.read.audio` now add argument `channel_first`

Fix:
- import `tensorneko_util` failed when pytorch is not installed

Features:
- Add `psnr` and `ssim` related metrics function to evaluate image and video
- For the metrics reduction option, now they're using `Reduction` enum.

Features:
- `crop_with_padding` to crop any location for a given image with zero padding.
- Refactor `Stream`. Now it supports lazy evaluation for all operators.
- To adapt ddp accelerator, add `distributed` argument for `NekoModel`.
- A `Timer` as decorator and context manager to count the runtime of a segment of code.
- `AutoEncoder` architecture in `tensorneko.arch`


Fix:
- Fix when starting web watcher server the program cannot find the directory
- Fix the tensorneko convolution layers that cannot be pickled in ddp.

Features:
- Move web watcher to `tensorneko_util` library.
- Refactor attention layers API
- Add `MaskedConv2d` for masked convolution layers from PixelCNN.
- Add `GatedConv` module from PixelCNN.
- Add `VectorQuantizer` layer from VQVAE.
- Add predefined architectures, `GAN`, `WGAN`, and `VQVAE`.

Fix:
- Remove `clean-fid` dependency.

Features:
- `MultiPlots` for easier displaying multiple plots in Jupyter Notebook.
- Merge namespace of `util.func` and `util.type` for both package

Fix:
- Refactor notebook display tools