[release/2.6] fix scaled matmul and test_float8_basics_cuda

[dnikolaev-amd]

This PR fixes:

• test_matmul_cuda.py::TestFP8MatmulCudaCUDA::test_float8_basics_cuda - AssertionError: RuntimeError not raised
• test_matmul_cuda.py::TestFP8MatmulCudaCUDA::test_scaled_mm_vs_emulated_row_wise_bfloat16_cuda - AssertionError: Tensor-likes are not close!

need to swap A_SCALE and B_SCALE descriptors data if use_rowwise like as HIPBLASLT_VEC_EXT

Fixes SWDEV-544098

[rocm-repo-management-api]

Jenkins build for efe26b4767ed81ed164dfdbd99473327691021d7 commit finished as FAILURE
Links: Blue Ocean view / Build artifacts

[jagadish-amd]

in PT 2.7, scales are swapped in https://github.com/ROCm/pytorch/blob/release/2.7/aten/src/ATen/native/cuda/Blas.cpp#L131
but in Pt 2.6, I do not see scales as parameter in https://github.com/ROCm/pytorch/blob/release/2.6/aten/src/ATen/native/cuda/Blas.cpp#L98
That might be the reason why the scale swaps are required further down the call in scaled_gemm in CUDABlas.cpp