BatchReindexLayer to shuffle, subsample, and replicate examples in a batch

[cdoersch]

This layer takes two blobs as input and uses the second blob to index the first blob along the first axis.

This layer is needed for the training procedure described in this paper, which first samples patches and extracts fc6-like features, before shuffling them around and pairing them up halfway through the network. @shelhamer agrees that this may be useful to others, hence this PR. It has documentation in comments and tests; this should be ready for review.

[jeffdonahue]

Thanks @cdoersch, this mostly looks good. For merge the main issues I see are related to the use of legacy 4D blob indexing in this layer (which is not specific to images). Specifically:

• uses of num() should be removed (change to shape(0))
• the shapes in the docs should be updated (e.g., first blob should be (N \times ...) and second blob should be 1D (N)(M))
• check should be added that the second blob is 1D (CHECK_EQ(1, bottom[1]->num_axes())) and the dimension of its only axis equals the dimension of the first axis of bottom[0] (again using bottom[0]->shape(0)) edit: sorry, I misinterpreted the layer's behavior with this last suggestion.

[jeffdonahue]

Thanks for the cleanup @cdoersch, LGTM.