Added all CPU to Docker GPU images for 'token_embd.weight' compatibility

[rudiservo]

Issue #12500, Cuda docker images are crashing with old GPU and maybe more recent ones because token_embd.weight is being processed by the CPU, since BMI2 was added this causes the program to crash due to compatibility issues.

It was recommended to add all CPU variants to all GPU images because it would benefit GPU images with CPU compatibility.

[slaren]

Looks good, but if any of these images are intended to be used on Arm (maybe Vulkan?), it would need the same logic as the CPU image to disable GGML_BACKEND_DL on Arm builds.

[rudiservo]

As far as I can tell only CPU images are built with ARM, all gpu images are built for AMD64.
If you want GPU builds for ARM, I can add them in a latter PR maybe? I don't have an ARM to test it with a GPU, so I will need someone to test it.

[aubinkure]

I was running into the same segfault running llama-quantize in the Docker container as these issues:
#11683
#12564
#11196

I traced it back to this PR. Any thoughts on how to fix it?

Separately, it seems like there's no CI on the Docker images? A few short tests would be quite helpful, I think.

[rudiservo]

@aubinkure how can you trace it back to this PR if this was more recent than the issues you refereed to.
you traced it forward m8, either this solves it or it doesn't, the only thing this does is to add CPU support to all GPU images.
If anything it's an issue with BMI2, so not this issue or PR.

[aubinkure]

You're right, it seems this PR isn't related to the issues I mentioned. That said, I double checked and this PR is definitely breaking quantization for me inside a Docker container. I'll open up a new issue with steps to reproduce.