Compute pairwise distances for all items in a dataset, for all layers

[sergeyk]

In the course of iterating over the data, Caffe can compute distances between feature representations of images at all layers in the convnet, and output them to file. This will enable exploration of convnet features for image retrieval, and make a nearest neighbor classifier trivial.

[kloudkl]

#243 has something to do with this requirement.

[shelhamer]

This seems like it would be done most simply through pycaffe or matcaffe (2) since all layers' data, diffs, and params are now exposed.