add contrib package

[piiswrong]

@precedenceguo @zhreshold
Moved extra operators to contrib. Had to fix lint, please check if I made any error.
BTW do you have unittests for these ops?

[mli]

if i understand correctly, we put ops in contrib if they are new and not well tested. once they are mature, we will consider to move them into core.

then do we really want to have contrib in mx.sym.contrib.xxx? if we move the ops,
then users need to update their codes.

[piiswrong]

contrib is for things that we don't give backward compatibility promise. When ops move to core interface/behavior may change

[zhreshold]

Does it mean all contrib ops are enabled by default?
I agree that some ops that are specific to certain tasks should be put somewhere else, but do you have a rule that what should be in and what should not? @piiswrong
BTW, I don't think backward compatibility is a big issue for those Ops metioned.

[piiswrong]

It's basically a namespace that indicates limited guarantees. volatile interface/behavior, untested, etc

[ijkguo]

I think it is better not to include them by default. Perhaps we can have a ENABLE_CONTRIB Makefile config to hide these operators since It could be confusing without examples.

Operators not core to deep learning are moved to this contrib package so that newcomers do not need to know them. For example, Correlation, BilinearSampler, GridGenerator, SpatialTransformer. If some operators get used frequently like BatchNorm, they could be moved to the main package.

@mli As to ROIPooling, smooth_l1, Proposal, MultiboxPrior, etc, I think these are special operators only for one purpose. Their usage is limited and one-time. It would not affect many users.

[piiswrong]

We can probably add an option to disable it in the future when compile time is too long. For now keeping them in contrib is fine.

[fullfanta]

@piiswrong

please fix the bug on faster-rcnn example.

At setup.py in python, packages contain 'mxnet.contrib'.

But at faster-rcnn example(https://github.com/dmlc/mxnet/blob/master/example/rcnn/rcnn/symbol/symbol_vgg.py), it uses mx.symbol.contrib.Proposal rather than mx.contrib.symbol.Proposal.