Make configurations consistent

[dcfidalgo]

Is your feature request related to a problem? Please describe.
There is a slight inconsistency when creating a PipelineConfig from scratch: for example the encoder in the CharFeatures or the feedforward in the TokenClassificationConfiguration have to be specified as python dicts instead of Seq2VecEncoderConfiguration and FeedForwardConfiguration, respectively (see this tutorial).

Describe the solution you'd like
Make it more consistent by either use *Configurations objects for everything, or take advantage of AllenNLP`s from_params machinery for everything. Personally i prefer the latter, but no strong opinion here.