Add EXLA to enable use of at least CPU-based ML

[lawik]

From working with Paulo Valente and Jonatan Klosko it seems we need to run from unreleased versions right now. Not sure what fix it is we need but that was the case when I tried to run Whisper.

[lawik]

I think this is all that's needed. I didn't build and run it. If you need me to, I can.

[fhunleth]

This is really amazing. Thanks for getting this started.

[fhunleth]

I rebased on main and updated to use hex.pm versions of everything now that they've been released. It seems really close, but still doesn't compile on CI. I'm able to build on MacOS with a manual change to the exla Makefile. That change shouldn't be needed on Linux, so I'm surprised CI didn't work, but will have to look into more later.

[lawik]

Let me know if you need me to try this on anything.

[fhunleth]

@lawik Could you check it? I think it works, but it's pretty easy to hang my RPi5 when using Bumblebee, so I'm not sure.

[lawik]

I'll give it a bash :)

[fhunleth]

This PR requires elixir-nx/nx#1543 to build on MacOS and on Linux machines that have CUDA tools installed. There are a couple ways this could be fixed, but that's why I haven't merged it yet.

[fhunleth]

I rebased again.