Fix border mode in 'Convolutional autoencoder'

[23pointsNorth]

By default border_mode='valid', thus there will be a tensor difference between the inputs and outputs.

Added , border_mode='same' as in the layer above within the model.

This fixes the code here.