[MX] Support mixed MXFP4/FP6/FP8 linear layer

[balancap]

Blackwell hardware natively supports any combination of MXFP4/FP6/FP8 in matmuls. See PTX and Cutlass documentation:

• https://docs.nvidia.com/cuda/parallel-thread-execution/#tcgen05-kind-shapes
• https://github.com/NVIDIA/cutlass/blob/main/tools/library/src/reference/block_scaled_gemm_mixed8bitsa.cu

According the MX paper, and macore generally the large quantization literature, there is advantages to use different bitwidth for weights, activations and gradients. It would be very useful in mx_mm and MXLinear to support this more general setting.

[balancap]

#1667 is adding mixed element dtypes support to mx_mm. A similar general interface could be added to MXLinear

class MXLinear(torch.nn.Linear):
    @classmethod
    @torch.no_grad()
    def from_float(cls, mod, in_elem_dtype, w_elem_dtype, grad_elem_dtype, block_size):
        ...

[jainapurva]

@vkuzo @drisspg

[balancap]

Closing the issue as #1667 has been merged