Dilated convolution support

[mdraw]

Dilated convolutions can be used to improve the context-awareness of CNNs. It would be interesting to compare dilated CNNs to the alternative multi-scale approach of the already supported U-Net-like architectures (example).

Implementing this feature is probably not much work, since it is directly implemented in Theano's convolution operation (depends on #4). This will require theano>=0.9.

I don't think we need to subclass elektronn2.neuromancer.Conv for this, but we can add this functionality directly to the Conv code like they did in Keras (see the dilation_rate parameter of the Conv*d constructors here).