Feature request: Depthwise and groupwise convolution

[ngphuoc]

These convolutions are increasingly popular in state of the art methods.

Related work done in Flux, NNlib, CuArrays: JuliaGPU/CuArrays.jl#523

[denizyuret]

Related to #407. Any recommended reading describing these operations?

[ngphuoc]

I missed that issue. Thanks. Will post there