refactor: separate dataframe and dataset

[maartenbreddels]

We're back to Dataset again 😄. An immutable mapping of name to column.

The idea is to abstract away the data part of a DataFrame, which gives us a few advantages.

• We can have lazy loading of data, and reading of multiple columns in 1 go
  • e.g. SQL from_sql and to_sql would be really helpful features in vaex #864
  • Parquet via Arrow Dataset API
• Dataset and thus DataFrames can be serialized (e.g. only store the hdf5 or arrow path, not serialize the data), so we can efficiently pickle/serialize/transport them with use of Dask/Ray/Mars.

This also starts with the idea of identifying data with a hash key to quickly compare data, which makes caching easier (e.g. also when using Dask), or when we want to cache complex operations (groupby).

We will not have subclasses of DataFrame anymore (except for DataFrameLocal and DataFrameRemote), all data specific
parts are done in the Dataset.

[maartenbreddels]

I can get it to work with both ray and dask, but dask crashes when doing distributed, not sure why. I think we want to ray/dask stuff on a different PR, but for now i keep this together. 2 commits in 1 PR would also be fine I think.