mx: make CUDA kernel for dim1 cast in mxfp8_cublas recipe

Summary:

As titled, this is the fastest option right now so should be default.

Test Plan:

```bash
pytest test/prototype/mx_formats/ -s -x
```

Reviewers:

Subscribers:

Tasks:

Tags:
ghstack-source-id: 52eb5cd
ghstack-comment-id: 3144899603
Pull-Request: #2661

Author: vkuzo

torchao/prototype/mx_formats/config.py

@@ -184,8 +184,10 @@ def from_recipe_name(
         if recipe_name is MXLinearRecipeName.MXFP8_EMULATED:
             return MXLinearConfig()
         elif recipe_name is MXLinearRecipeName.MXFP8_CUBLAS:
-            # TODO(future PR): default to CUDA dim1 kernel
-            return MXLinearConfig(gemm_kernel_choice=MXGemmKernelChoice.CUBLAS)
+            return MXLinearConfig(
+                gemm_kernel_choice=MXGemmKernelChoice.CUBLAS,
+                mxfp8_cast_kernel_choice=MXFP8Dim1CastKernelChoice.CUDA,
+            )
         elif recipe_name is MXLinearRecipeName.MXFP8_CUBLAS_RCEIL:
             return MXLinearConfig(
                 gemm_kernel_choice=MXGemmKernelChoice.CUBLAS,