Is your feature request related to a problem? Please describe.

Disentangle scale swizzling from GEMM.

Describe the solution you'd like

• Add logic in C++ tensor class whether row-wise/col-wise scales are swizzled or not
• Add logic in PyTorch quantized tensors whether row-wise/col-wise scales are swizzled
• Add hints in quantizer whether pre-swizzling is helpful

Describe alternatives you've considered

Additional context