SQUASH layer -- takes 1 or more inputs and produces no output

[jeffdonahue]

I used this to suppress outputs that I didn't want printed during training. Perhaps we shouldn't have a layer for this, I don't know. But here's a PR.

[sguada]

I guess it can be useful if one want to discard some blobs and not appear as output_blobs. However why should they compute backward?
How about SINK layer for name?

[shelhamer]

To join the naming party, how about SILENCE? Instead of defining it as a layer in the prototxt what about adding a silence field to the layer message that takes a top blob name and instantiates a silencing layer as directed.

I'd give this a vote for convenience, since depending on phase or purpose one might want different output blobs.

[jeffdonahue]

I thought the SQUASH name would probably see some backlash; I'll change it to SILENCE when I get a chance. There is normally no need to compute backward, but I put the gradients in because they're very obvious -- if #686 is merged, the net will be able to figure out that it doesn't need to compute the backward pass of the SQUASH/SILENCE layer (as it has no top blobs, so it can neither produce a loss, nor send input to later layers that do produce a loss) so with that change the Backward implementation is only included for completeness -- it would only actually be executed if force_backward is enabled or something.

[longjon]

This is rebased, renamed, and ready-to-go, pending Travis.

[jeffdonahue]

Sweet, thanks for fixing this up! I can't figure out why the Travis build is erroring though...

[longjon]

It's because I didn't add a STUB_GPU definition for CPU_ONLY mode. Hopefully I've fixed it now.