[WIP][Do not merge][video] CSN networks

[bjuncek]

Initial implementation of the Channel Separated ConvNets as described in https://arxiv.org/abs/1904.02811, with belonging code and models.

Please do not merge this PR before 1 and 2 are not merged into their respective masters.

[1] Slightly changes the video model API - this has been tested with and without the changes there and it works, but for the sake of consistency, let's wait until they all look the same.
[2] There is no efficient implementation of the 3D depthwise convolution, and until there is, the FLOPs saved by using depth-wise channel-separated convolutions is meaningless given the inefficiencies in the kernel.

---

Discussion point:

Which pre-trained model do we want to provide with them kinetics as everywhere else, or a more suitable one such as Sports1M or URU?

[fmassa]

@bjuncek

Which pre-trained model do we want to provide with them kinetics as everywhere else, or a more suitable one such as Sports1M or URU?

Let's keep using Kinetics-400, as that's what the other models use already, and it is also referenced in the CSN paper.