Add groups support to ResNet

[fmassa]

This adds the necessary building blocks to add support for ResNeXt models in torchvision.

One thing worth discussion: should we let the user specify arbitrary number of channels in each layer (as is currently supported in BaseResNet), or just stick with the options that are necessary to cover ResNet / ResNeXt?

[codecov-io]

Codecov Report

Merging #822 into master will decrease coverage by 0.04%.
The diff coverage is 92%.

@@            Coverage Diff            @@
##           master    #822      +/-   ##
=========================================
- Coverage   52.45%   52.4%   -0.05%     
=========================================
  Files          35      35              
  Lines        3401    3408       +7     
  Branches      545     547       +2     
=========================================
+ Hits         1784    1786       +2     
- Misses       1485    1488       +3     
- Partials      132     134       +2

Impacted Files	Coverage Δ
torchvision/models/resnet.py	83.1% <92%> (-0.58%)	⬇️
torchvision/transforms/transforms.py	82.84% <0%> (-0.68%)	⬇️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update a61803f...0e1e363. Read the comment docs.

[ekagra-ranjan]

ResNets and its variants find their place as a building block in many new models. So providing a flexibility in choosing the number of channels (besides choosing the layer configuration) would give more power and ease to users while constructing their own custom models. And since there is a relation between the number of channels and groups, the implementation is quite neat and will remind the user (experimenting the source code) about the relation proposed by the authors.

[fmassa]

@ekagra-ranjan just to make sure I understand, were you advocating for passing the layers argument (as it was before, in the BaseResNet) or just the width_per_group ?

[ekagra-ranjan]

I am supporting the BaseResNet implementation that you implemented earlier in the PR which has planes and layers arguments.

[fmassa]

I'll be merging this as is for now to unblock a few other things, and I can add back BaseResNet in a follow-up implementation.