stacks: consolidate ml pipelines

[alalazo]

Extracted from #3175

This commit aggregates the ml-linux pipelines per architecture. As such:

• It reduces the number of pipelines from 5 to 2
• It removes a lot of redundant, duplicate, builds from pipelines

Closes spack/spack#32893

[alalazo]

@adamjstewart

[adamjstewart]

If you want you can also reformat darwin the same way with shared definitions. We could build both cpu and mps for darwin. PyTorch is the only package with an MPS variant, so it won't add too many builds.

[alalazo]

If you want you can also reformat darwin the same way with shared definitions. We could build both cpu and mps for darwin. PyTorch is the only package with an MPS variant, so it won't add too many builds.

Let's do it in another PR. I'd like to keep this PR just focused on consolidating the existing ml-linux pipelines.