Change features vs schema logic

[lhoestq]

New logic for nlp.Features in datasets

Previously, it was confusing to have features and pyarrow's schema in nlp.Dataset.
However features is supposed to be the front-facing object to define the different fields of a dataset, while schema is only used to write arrow files.

Changes:

• Remove schema field in nlp.Dataset
• Make features the source of truth to read/write examples
• features can no longer be None in nlp.Dataset
• Update features after each dataset transform such as nlp.Dataset.map

Todo: change the tests to take these changes into account

[lhoestq]

I had to make SplitDict serializable to be able to copy DatasetInfo objects properly.
Serialization was also asked in #389

[thomwolf]

Really cool!

[thomwolf]

One thing I forgot to say here, is that we also want to use the features arguments of load_dataset (which goes in the builder’s config) to override the default features of a dataset script.