[Bug] depthwise convolution implementation seems wrong

[roytseng-tw]

https://github.com/rwightman/pytorch-image-models/blob/9a25fdf3ad0414b4d66da443fe60ae0aa14edc84/timm/models/layers/create_conv2d.py#L25
The value of groups should be set to in_channels instead of out_channels.

[rwightman]

@roytseng-tw in this use case for this layer it doesn't matter, there is no depth multiplier, without a depth multiplier in_channels == out_channels .... do you have a network were it does matter?

Also see #241

[roytseng-tw]

I just want to remind that if someone uses this API with in_channels != out_channels, then an error will occur.