PULPOpen FP32 Conv2D (regular and DW) kernel improvements

[diaconuccalin]

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

• Channel-wise tiling currently not supported by Conv2D DW
• Group parameter currently not supported by regular Conv2D kernel

Describe the solution you'd like

• Analyze whether it would make sense to merge the 2 kernels (regular and DW) into a single one, as it is handled in PyTorch, or keep them separate and make the necessary changes in the kernels, tile constraints, and bindings

Describe alternatives you've considered

• Potentially extra sub-feature: analyze and, if suitable, implement in-place DW conv from tinyengine