making input and output shapes explicit to support nonlinear models

Created by: yiiyama

Replacing the variable current_shape with more explicit input_shapes and output_shapes variables. This adds the flexibility needed for some layers that take multiple input tensors. The output_shape of each layer is inserted to the output_shapes dict at line 432 (for skip layers in line 167-170) with the name of the layer as the key. The inbound_nodes parameter of the layer config in JSON is given in terms of the layer names, so we can look up the shapes from this dict to set the input_shapes list (lines 154-157).