Removed inline mode, changed generation of execute and python execute and implemented beam search for both python and c++

Added possibility to save the visual attention for Show, attend and tell architecture as an image. Adjusted axis parameter in softmax layer to ignore batch size. Fixed a problem with some states not being initialized correctly in C++ code due to previous changes to the inline mode

Added BroadcastAdd and Multiply layer. Renamed Multiply layer to BroadcastMultiply. Added axis parameter to Softmax layer. Added SoftmaxCrossEntropyLossIgnoreIndices to SupervisedTrainer for now, in order to ignore certain words/indices for loss computation in NLP applications

fixed various bugs, in particular to allow integer outputs in C++ code generation and to fix shape errors for LSTM inputs