Fix/symmetric contraction groupmm bindings

[asglover]

Re-enable symmetric contraction / groupmm bindings and added basic pytest testing.

[vbharadwaj-bk]

Overall, looks great. Couple of things:

1. RocM, I'm assuming hasn't been tested. Are you basically adding back in code that was taken out? In which case, no problem. Otherwise, I'm a bit wary.
2. Have you run the test suite / installation on google collab with an NVIDIA GPU, or on a platform of your choice? I would do that, since this requires changes to the low-level extension.
3. No problem adding cuBLAS back to the CMakeLists.txt for the stable extension, since someone wants this. I was really hoping to not do this; without it, we could package up a prebuilt wheel and avoid compilation on the user side entirely. Now, we can't do that, and worse, the user's cuBLAS should be ABI compatible with PyTorch's installed version. At some point, we should address this.
4. Future work (not in this commit): elimination of the groupMM custom class. You can switch to a caching scheme similar to the regular OEQ kernels for the cuBLASLt state. This will enable you to compose this with torch.compile (and, incidentally, you should tell the user that this is not compatible with torch.compile (yet). Then you can add a symmetric contraction test in the export_test.py file.