FP4/UINT4 packing order

[hann-wang]

Both ONNX and OCP MXFP requires the first element in LSB.

• https://onnx.ai/onnx/technical/float4.html
• https://onnx.ai/onnx/technical/int4.html
• https://triton-lang.org/main/python-api/generated/triton.language.dot_scaled.html#triton.language.dot_scaled

But torchao packs the second element in LSB:
https://github.com/pytorch/ao/blob/main/torchao/prototype/mx_formats/kernels.py#L761

[vkuzo]

yes, we pack the second element in MSB in PyTorch core (https://github.com/pytorch/pytorch/pull/148791/files) and we need to update torchao to match.