Allow ReXNet to use other activations

[chuong98]

Hi, in the current version:
pytorch-image-models/timm/models/rexnet.py
the activation is hard coded as "swish". It will be more flexible to use other activations, such as ReLU.
Thank you.

[rwightman]

@chuong98 I thought about this while working on it but didn't come up with a great idea for the arg interface so left it for now, I want to get a big validation run started with all models (incl new ones and weights across all of the test sets) ASAP so I can get the features branch merged.

One thing I didn't decide on was what to call the two different activations, I use 'act_layer' as an arg when changing the act is supproted, wasn't sure what to call the swish vs relu ... act_layer and smooth_act_layer, or maybe exp_act_layer (exp for expansion) with the idea being you'd normally set act_layer to relu6 or relu and smooth_act_layer to something like swish, mish, elu, gelu....

You could set the acts all the same, but that is one of the key differences of this model vs from other MBConv nets like EfficientNet, MobileNetv2, etc

[rwightman]

Added act and dw_act args in #475