Fix broken DeconvolutionLayer GPU backward caused by ND conv typo

[longjon]

#2049 broke deconvolution GPU backward with a typo (https://github.com/BVLC/caffe/pull/2049/files#diff-5911f1d28796ada5f80449e8bd05d270R54 omits the n * coefficient). This PR fixes.

This should really really be covered by the tests, but this PR is just the fix for master.

[jeffdonahue]

Yikes! Thanks for the fix. Looking quickly (since I want to merge ASAP) I don't understand why the buggy version passes the current gradient checks...it seems like they shouldn't pass unless we were only testing the n == 1 case, which we're not -- the tests have n == 2.

[shelhamer]

@jeffdonahue @longjon It is clearly good that the indexing of the top diff is fixed so that deconv is correct in all cases, but oddly enough the code was correct as it was when doing backprop to both params and bottoms according to param_propagate_down and propagate_down. This case hinges on passing this->param_propagate_down_[0] as the skip_im2col arg of forward_gpu_gemm: in this way the weight gradient call first im2cols the right part of the top diff according to top_diff + n * this->top_dim_ and then the bottom gradient call ignores its own input arg in favor of the cached im2col result. As the gradient checker backprops to weights and bottom(s) in CheckGradientExhaustive() this condition always holds during the tests.

All-in-all this was subtle and weird but not completely broken. However, we should consider test coverage of different param_propagate_down and propagate_down settings since these could have caught this and the concat partial backprop problem #2972 .