ci: Fixes for pre-built wheels

[ko3n1g]

This brings support for NGC PyTorch containers and removes cuda11 since it doesn't seem to be compatible anymore

[ko3n1g]

Here's a passing run: https://github.com/ko3n1g/DeepGEMM/actions/runs/18443482987

[ko3n1g]

After merging this one, can we create and push a new tag @LyricZhao ?

[LyricZhao]

After merging this one, can we create and push a new tag @LyricZhao ?

Sure! Can you bring Python 3.8 back?

[ko3n1g]

Sure! Can you bring Python 3.8 back?

which cuda and torch version do you expect/ want to support with py3.8?

[LyricZhao]

How about CUDA 12.3 + PyTorch 2.1.0?

[ko3n1g]

Okay, looks like we've got it now. Here's the job that also builds a wheel for Py3.8/Cu12.3,PyT2.1.0: https://github.com/ko3n1g/DeepGEMM/actions/runs/18457600231/job/52581827327

Since support for Py3.8 is quite limited I didn't include a vast grid but just this singular configuration

[ko3n1g]

I think we need to use cuda11.8 instead: https://github.com/pytorch/pytorch/blob/main/RELEASE.md#release-compatibility-matrix

[ko3n1g]

@LyricZhao Finally got it working with cuda12.1: https://github.com/ko3n1g/DeepGEMM/actions/runs/18465627532/job/52608861167

[LyricZhao]

LGTM, ready for merge now?

[ko3n1g]

Yep!