WIP Cupy integration

[dmenig]

I did this to change torch tensor storage of decoded frames to cupy.

This is because the memory footprint of cupy is much lower than pytorch. Since I plan on using this in a multiprocessed env, I needed the footprint on the GPU per process to be low. This should allow me to use ~50 processes on a 16GB VRAM GPU, which is much better than about 15 with pytorch.

I wasn't really planning on merging this, and doing the work to allow for storage nature choice with a flag or whatever. I planned on just making it run for my specific use in particular (hence many deleted functionnalities). But I think it's still good that people coming here know that cupy integration is possible.

Also I tried jax, but the imutable nature of the jax arrays prevented me from doing anything.

Weirdly, I noticed I need to import cupy before torch, or else cupy fails to import !

Note that torch / cupy interoperability is alive and well.

[anibali]

That's pretty cool, thanks for making your code visible here! It will definitely save others time if they have the same idea of using cupy+tvl.