Move on to Parquet format

[mikegerber]

33 and related problems look like we finally move to the Parquet format - sticking to an old Pandas version to produce a "stable pickle" does not seem maintainable anymore.

• [x] Review "set" columns like classification-ZVDD
• [x] Review dtypes

[mikegerber]

• Set columns become "arrays" now, due to conversion to Parquet and back
• dtypes are as before

→ Defering improvement on this to #37