[RFC] feat: more precise qbinary handling in tract core with unaligned entry

[JulienBalianSonos]

Change ways binary ops are done for math with quantized tensors, by allowing unaligned qparams in input/output tensors.

This change allows increasing precision of quantization and reduce casting/alignment cost, but we now need to handle the qparams of output tensor as function input.

We need to handle all ops implied by bin_to_super_type macro already supporting QU8, that may use unaligned qparams quantization.

• elementwise mul
• div

in next cases we directly apply dequant to f32 and requant to targeted qtype for now

• add
• sub
• min
• max (here we set a small optim to make it slightly faster for ReLU case)

we propose to leave aside from this PR: rem, and all "bool" or "bit" operations