This adds ScalarLayer, which takes two Blobs and (in effect) multiplies them elementwise, after broadcasting the axes of the second Blob to match the first as necessary.

For example, if bottom[0] has shape (2, 3, 4, 5) and bottom[1] has shape (3, 4) and axis == 1, then the computation of this layer is equivalent to reshaping bottom[1] to (1, 3, 4, 1), then tiling it to (2, 3, 4, 5), then multiplying the result elementwise with bottom[0].

In the most general case, Backward to bottom[1] is accomplished with elementwise multiplication followed by 2 gemvs. For special cases (when bottom[1]'s shape corresponds to the beginning or end of bottom[0]'s shape, e.g. if it were instead shape (2, 3) and axis == 0 or shape (4, 5) with axis == 2) one or both of the gemvs is skipped (or replaced with a dot product).

My use case for this comes from #2033 -- I am replacing the hacky coeff_blob I added to Eltwise to perform the binary multiplications with this layer. It could also replace the channel-wise scalar in PReLU (I think this backward implementation is faster), or be used to learn a channel-wise scalar after batch normalization.

Thanks to @longjon for the name for this layer and initial implementation of a previous version.