feat: initial amd gpu support with rocm 7.1

[pyqlsa]

Summary

Intended to address #27 by adding ROCm support for AMD GPUs.

Adds

• ROCm support for AMD GPUs for the ComfyUI package (via ROCm 7.1); tested gfx1100 with a 7900XTX, and a couple of the default text-to-image and image-to-image template workflows appear to be working as one would expect; there are a few gotchas with ROCm support in general (e.g. xformers), but there look to be known workarounds with tunables nominally available with ComfyUI; CPU architecture compat is only x86_64.
• ROCm compatible app so we can nix run .#rocm; (tested gfx1100 with a 7900XTX); CPU architecture compat is only x86_64.
• ROCm compatible container image so we can nix build .#dockerImageROCm; (tested gfx1100 with a 7900XTX running with podman); running with docker not yet tested). CPU architecture compat is only x86_64.
• Extended options for the ComfyUI nix module to support configuration to leverage the rocm variant (tested on my own system configuration).
• Updated the CI workflow to build ROCm artifacts (largely just copying the pattern for CUDA artifacts).
• README updates in line with the above.
• Another dev shell, just for ROCm, which pulls in the python environment that is ultimately passed to ComfyUI; it was helpful for exploration and a bit of debug.

Changes

• Changed the cudaSupport toggle to a gpuSupport enum; users (and internal plumbing) now select between cuda, rocm, and none (for cpu-only).
• Updated the container section of the README to differentiate between the apps that build and load the container vs. just building the container and manually loading with your tooling of choice (either docker or podman).

Observations/Comments

• Is it worth it to create apps for podman like there are for docker?
• Is it worth it to also try a spin with ROCm 6.x? (Maybe not, looks like unstable is already on 7.1).
• I didn't need to pass autoPatchelfIgnoreMissingDeps anything for sox or ffmpeg; I'm not quite sure why that was required for CUDA and not for ROCm.
• Something in ComfyUI shells out to rocminfo on launch, and I can't figure out yet what it is; I tried adding it as buildInput and propogatedBuildInput in torch, and in the comfyui derivation, but nothing seemed to stick; even without it, ComfyUI starts just fine, but writes some warnings to the logs that it couldn't find rocminfo, but it was able to detect ROCm and GPUs anyways; I threw rocminfo into the container and the systemd unit path to make those warnings go away, regardless.
• Couldn't seem to figure out how to ignore the test_optim test for timm, to I tried just hucking setuptools in there, and it seemed to work; based on the surrounding comment and ignoring timm's test_kron test, I'm not sure if my choice is an anti-pattern or not.
• There are still some TODO and XXX comments in there that overlap with some of these thoughts; I'm happy to clean those up before a potential merge, but they're in there for now in case they prompt any specific feedback.

[jamesbrink]

Thanks @pyqlsa .. sorry i missed this, I cherry picked your commits as i started a version upgrade before I saw this PR. please re-open issue or PR if new release does not work.

#30

[jamesbrink]

Your commits were cherry-picked into #30 (preserving your authorship) with bug fixes applied on top. Merged and released as v0.14.2. Thank you for the ROCm implementation!

[pyqlsa]

No worries at all. I'm back to the keyboard and will check this all out. Thank you @jamesbrink !