Add parameter layer for learning any bottom

[longjon]

(This is a minimal step in the direction of #1474. From discussion with @jeffdonahue.)

This layer simply holds a parameter blob of user-defined shape, and shares it as its single top.

This is useful if you want to learn disconnected bottoms in a net. In theory, all parameters could be handled this way, which would be a full realization of #1474. It is, however, not clear that that's the right thing to do from a user interface perspective: params would lose their semantic distinction, and their sizes would end up being double-specified since current layers already compute them.

Whether or not this ends up being on the solution path, I wanted to go ahead and throw up this PR since I'm already making use of it with other, future PRs.

[jeffdonahue]

Cool, this looks good to me -- feel free to merge as you see fit. (I did realize after our discussion, that this couldn't just be used with an Eltwise SUM/PROD to emulate a hypothetical BiasLayer or DiagInnerProductLayer since the params would need to be tiled to match the number of instances. I should PR my TileLayer which would then let you do that sort of thing with this... Edit: see TileLayer in #2083)

[longjon]

This has now been rebased, as I'm continuing to use it in some of my models. I'll plan to merge soon per @jeffdonahue's prior approval, unless there are any further comments (in particular, if master has changed in ways that affect this PR -- I can't think of any).

[jeffdonahue]

still LGTM

[shelhamer]

Fine by me. This does come up from time to time so let's merge.

[longjon]

Okay, thanks for the eyes, merging this simple layer which I've made quite a bit of use of myself.