Enable builds without direct torch.cuda availability and support sm89 / sm90.

[wanderingai]

This PR allows for monarch_cuda kernel to be built for computes sm80, sm89, and sm90, which includes the following GPUs:

• A100
• H100
• L40
• RTX 6000 Ada

Additionally the setup.py is updated to enable builds based on nvcc availability but without direct torch.cuda availability for flexible builds.

Update:

The compiler flags have been abstracted to support both PTX and SASS builds while defaulting to the original Ampere-based PTX-only build i.e. -gencode=arch=compute_80,code=compute_80.

Successfully tested by building a docker image and running tests under tests/:

---

---

[DanFu09]

It looks like this PR is introducing some race conditions - when I install using this branch, some tests fail:

pytest -s -q tests/test_flashfftconv.py
Running 1120 items in this shard
......................................................................................................................................................................................................................
F.....................................................................................................................................................................................................................
................................................................F.F...F...............................................................................................................................................
......................................................................................................................................................................................................................
......................................................................................................................................................................................................................
..................................................

[michaelfeil]

@wanderingai Love this PR, its a improvement from the previous setup.py. Can this be merged, worst case with no sm_90 flags by default?