addition of MultiBindStackConv2D class and addition to pipeline.

[ilibarra]

• Implement class MultiBindStackConv2D, that inherits from MultiBind, and uses a stack of Conv2Ds instead of conv_mono and conv_di independently.
• update parameter kernels to a list of lists, indicating (n_layers, size_layer). e.g [[0, -1], [3, 15], [4, 20]] means that the first kernels is empty (intercept), the second one has three Cond2Ds stacked, and the second one has 4 Conv2Ds stacked. The second parameter indicates a the max width for the first layer, and in subsequent layer the output goes down by one value for each layer. E.g. 15, 14, 13 and 20, 19, 18, 17 for layer two.
• the forward function needs to be updated to consider this case and be able to optimize.
• fit_model.py in mubind-pipeline needs to recognize as input a model class, chosen between MultiBind, MultiBindStackConv2D, to or another, and be executed as independent jobs, and output by directory {tf_name}/{fit_model}/{MultiBind} or {tf_name}/{fit_model}/{MultiBindStackConv2D}