Add TensorRT models to work

[My12123]

• Add TensorRT models for operation, the ability to run TensorRT models
• TensorRT Converter

[nagadomi]

On Linux, torch.compile() is supported. torch.compile should support the TensorRT backend in the future.

To use it, specify --compile option for waifu2x.web. In GUI/CLI, it is automatically used when the input is video.
Note that compiling the model can be very time consuming. Also, it is not yet supported for Windows.

Also, ONNX models can use the TensorRT backend. However, only JavaScript implementation(unlimited:waifu2x) currently use onnxruntime.

[nagadomi]

I have confirmed that upconv_7, cunet works with TensorRT backend. However, it is slower than with the default backend and no compiling.
For swin_unet*, it requires PyTorch 2.1, but TensorRT does not yet support PyTorch 2.1.

[alanzchen]

Looks like iw3 would also benefit from https://github.com/spacewalk01/depth-anything-tensorrt !

[nagadomi]

I have confirmed that using torch.compile with PyTorch 2.2 can speed up the waifu2x models almost 2x.
So I do not think we should stick to TensorRT.

For iw3's DepthAnything model, I have optimized it with mini-batch and fp16, so it should already be 2x faster than the official video demo.
(Probably the same approach as LiheYoung/Depth-Anything#49 (comment) )
I have not tried torch.compile for the DepthAnything model yet.
The one bad thing about torch.compile is that it does not work on Windows.