Support for extracting conv layer / spatial features for ResNets (for attention-based models)

[abhshkdz]

trafficstars

[abhshkdz]

Needs a command-line argument (like the VGG models) and corresponding model:remove() calls here: https://github.com/batra-mlp-lab/visdial/blob/master/data/prepro_img_resnet.lua#L37