Tests and fixes to new data api

[DeNeutoy]

In summary this PR does the following few things:

• Moves sort_by_padding from the Dataset to the DataGenerator
• Adds tests for : Vocabulary, Dataset, all concrete Fields, Embeddings, TokenIndexers, DatasetReaders(The reader ones could be more comprehensive, probably).
• Roughly fixes the DataGenerator to work with the new Dataset class; I haven't fixed the tests for this, because MattG wrote them and they rely on some FakeInstances, which I'm not exactly sure how to translate to the new api.
• Adds getter methods to the Fields, Instance and Dataset classes, called things like fields() for the Instance class, and tokens() for the TextField etc. I think it's important to be able to interact with the different classes so that writing functions for mapping predictions back to text etc is easy.

[matt-peters]

This looks great to me, from my somewhat quick look 👍 . I didn't see any big API changes in this, just testing and cleanup, right? It's nice to have usage examples in the test suite, thank you for adding them!