Add dropout

[hughperkins]

• dropout is a vital function, not currently implemented. pretty vital for any self-respecting convolutional neural network really ;-)
• it doesnt need to be added to all propagate and backprop implementations, but probably at least to:
  • Propagate1: generic gpu-based forward propagation layer
  • BackpropErrorsv2Naive: generic gpu-based backprop of error gradient to previous layer
  • BackpropWeights2Naive: generic gpu-based backprop of error gradient onto weights of same layer
• you'd need to also add the Dropout option to ConvolutionalMaker class
• and you'd need to make all other implementations (Propagate2 etc...) throw a runtime_error, in their constructor, if dropout is requested in the passed-in maker object

[hughperkins]

Following discussion with Tambet, implementing this as a separate layer. See https://github.com/hughperkins/DeepCL/tree/dropout

[hughperkins]

Done :-)