Fix per_channel quant axis for depthwise conv2d weights and bias (#903)

Summary:
Pull Request resolved: https://github.com/pytorch/executorch/pull/903

Rationale,
* PyTorch, weights = [out_channels, in_channels/group, kernel_h, kernel_w], per_channel quant axis = 0
* XNNPACK, weights = [in_channels/group, kernel_h, kernel_w, out_channels], per_channel quant axis = 3

This diff fixes the axis value (i.e. weight dim) and convert from 0 --> 3 before passing it on to XNNPACK just like we do weights already

Reviewed By: kimishpatel

Differential Revision: D50195930

fbshipit-source-id: cfc7b5a42a94f3d227462233582caa58c870018c