v0.2-p0

• Im2col convolution during prediction became faster, since some calculations now already happen already when loading a model.
• Support for non-im2col convolution was dropped.