Add Support for Providing IO Size for Primitives

[proska]

Some primitives like "d3m.primitives.common_primitives.FeedForwardNeuralNet" require the Input and output size to be specified as hyper-parameters. Currently, our system does not infer the data size during pipeline generation in "to_pipeline" method which makes adding support for this particular set of primitives a little challenging.
At this version of the code, we change configuration points to pipeline solely based on pipeline topological information. It works well with sklearn primitives as they internally infer the IO size during the fit. However, in other primitives, the behavior is not consistent.
We have to either completely ignore these primitives or evaluate the size of primitive outputs during pipeline execution.