Expand publish matrix for PyTorch 2.11 and 2.12

[qthequartermasterman]

This updates the release workflows to publish wheels for PyTorch 2.11 and 2.12, and aligns the reusable build workflow with the CUDA wheel variants those PyTorch releases actually publish. The matrix now includes the new supported combinations while excluding unsupported torch/CUDA pairs.

• Publish matrix

  • Adds 2.11.0 and 2.12.0 to publish.yaml
  • Adds CUDA 13.2.0 for the PyTorch 2.12 wheel line

• Compatibility rules

  • Excludes 11.8 for PyTorch 2.11 and 2.12
  • Excludes 13.2 for all versions before 2.12
  • Extends the existing cxx11_abi: FALSE exclusion to 2.11+, matching the newer pip wheel behavior

• Reusable build workflow

  • Extends _build.yml CUDA wheel selection logic:
    • PyTorch 2.11 → cu126, cu128, cu130
    • PyTorch 2.12 → cu126, cu130, cu132

torch-version: ["2.6.0", "2.7.1", "2.8.0", "2.9.1", "2.10.0", "2.11.0", "2.12.0"]
cuda-version: ["11.8.0", "12.9.1", "13.0.1", "13.2.0"]

[qthequartermasterman]

Hi @tridao and @ko3n1g! Are there any other spots in the workflow that would need to be updated to accommodate building wheels for these new python versions?

I have projects that use this kernel, but need to update to newer versions of PyTorch. My CI jobs are having to build from source over and over again. If the change to build official wheels is truly this simple, it would benefit not only me, but others I'm sure.

I believe this fixes #103.

Thanks!