Multiple outputs and restructured ParallelSourcePlugin

[JelleAalbers]

This allows strax plugins to have multiple outputs. If you want to return incompatible kinds of rows, e.g. both records and hits, or peaks and some metadata for each chunk, multi-output support is essential.

To return multiple outputs from a plugin:

• The provides tuple should have multiple elements, listing the provided data type names
• The dtype and data_kind attributes should be dictionaries mapping data type names to dtypes and data kind names, respectively. The values of these dicts can be specified in the same way as the entire attribute would be for a single-output plugin
• The compute method must return a dictionary mapping data types to results (structured numpy arrays or field/array dictionaries).
  For more info, see the updated docs, and tests/test_multi_output.py provides an example.

This also rewrites ParallelSourcePlugin. We should no longer subclass these plugins directy (as e.g. straxen's DAQReader does); instead, they are automatically created by strax when a parallel='process' plugin is followed by other plugins that can become inlined. Again, further description is in the docs. This should speed up processing from records onwards on machines with many cores, since records is then picked up by ParallelSourcePlugin just like raw_records usually is. See also the new docs and #150.