ROCm in WSL

[HPPinata]

This is a first draft. Feedback on the clarity of descriptions, level of detail and questions left open is welcome.

The WSL-Patch needs a bit more testing to make sure it runs reliably.
flash_attn integration is also flaky as of now.

Memory usage, as usual on Linux systems is suboptimal, WSL memory overhead makes this even more obvious.

[HPPinata]

Consider merging #347 before this, since this PR has been tested on top of it.

[tazlin]

I'm OK with this as is; the functional change to the scripts should only affect WSL and the doc changes can be iterated on if needed. I can't step-by-step confirm that it works as reported, as I don't have access to an AMD machine, but it looks reasonable and I'd rather merge now so the git history doesn't keep diverging.

[HPPinata]

I've tested a blind copy/paste of the commands from the wsl unregister stage to getting a worker up and running (with the only exceptions being a copy/paste of the bridgeData.yaml and models directory to avoid redownload) so it should work (Tested on Windows 11, with the 10.2024 drivers and a supported RDNA3 Card).
Any compatibility issues beyond that will need to be reported by users, I think that should be fine for something explicitly stated to be experimental.

Initially activating WSL is hard to test if it has been installed before, but if we get any questions around that it's just a Readme change to correct. The library patch should continue to work unless the "recommended" way to deal with this changes upstream or there are incompatible Driver/Kernel upgrades.

[tazlin]

Agreed on all.

Thanks again for your work on this and all your other initiatives.