[BUG] cutlass-python does not work on H100 for CUDA 11.8

Describe the bug
The following sample program fails with the latest version of cutlass:

import cutlass
import numpy as np

plan = cutlass.op.Gemm(element=np.float16, layout=cutlass.LayoutType.RowMajor)
A, B, C, D = [np.ones((1024, 1024), dtype=np.float16) for i in range(4)]
plan.run(A, B, C, D)

The error given is as follows:

Error Message:
nvcc fatal   : Value 'sm_90a' is not defined for option 'gpu-architecture'

Steps/Code to reproduce bug

• Download the latest version of cutlass on a H100 node that is configured with CUDA 11.8.
• Build cutlass by following the installation instructions and (optionally) specify -DCUTLASS_NVCC_ARCHS="90"
• Install the python library by running pip install . in the root directory of cutlass.
• Run the example program.

Expected behavior
It looks like cutlass' python wrapper expects all H100 machines to want to run with CUDA 12+. I see, for instance, the following:

https://github.com/NVIDIA/cutlass/blob/3a8c01a18b24c35b216922481ac762496720a99d/python/cutlass/emit/pytorch.py#L691-L694

But that isn't necessarily the case.

Start with python/cutlass/emit/pytorch.py around lines 691-694 and reproduce the issue using the provided GEMM sample after building with CUDA 11.8 on an H100. Trace how the GPU architecture is selected, then verify that installing and running the Python example no longer produces the unsupported sm_90a error under CUDA 11.8.