cuDNN frontend grouped GEMM fusion API is restricted to PyTorch

[Anerudhan]

The cuDNN frontend has added grouped GEMM fusions, such as grouped GEMM + swiglu + MXFP8 quantize. However, these are only accessible through the public API here, which depends on torch and torch Tensors.
https://github.com/NVIDIA/cudnn-frontend/blob/97f6cb3b88cacff507cca1280db5650a457d92b3/python/cudnn/grouped_gemm/grouped_gemm_swiglu/api.py

Since this depends on torch Tensors, we cannot call this from JAX/XLA. So we’d like one of two options:

1. CuTeDSL kernel is made public directly and a stable direct CuTeDSL kernel interface is committed to from cuDNN frontend team. JAX can call CuTeDSL kernels directly, it’s the public API torch Tensor wrapper that is problematic for us.
2. Expose this functionality through the C++ cuDNN frontend API. We can call the C++ cuDNN frontend API from JAX/XLA, but not the cuDNN frontend Python API as it depends on torch Tensors