New versions require protobuf >= 5 which we don't support

* Different versions of as_of
* date_rage
* row_range
* Different column slice size

…interface

… one for cols

…olumns

* Use shared_ptr for the pipeline context. It's needed because when
  queries are created a lambda captures the pointer and then the lambda
  object is returned. We want it to be copied so that the lambda is an
  owner.
* Add comments to pipeline fields (selected_columns_ and
  overal_column_bitset_)
* Fix pipeline context swap as it did not copy the overal column bitset

Mark the correct columns both in the selected_columns and overall_column_bitset.
After that the dataframe is read as usual and returned. The python layer
needs additional work since it cannot denormalize it (probably because
there is only one index column with no data columns).

TODO:
* The C++ is probably doing more than needed
* Python denomalization

The index columns are always the first columns in the stream descriptor.
Thus we can check the count of the index columns and add the first n
columns to the selected set, skipping the procedures of creating hash
sets and checking if a field name is in the set.

…filter

…the rowcount

…s or recursive normalizers when trying to read index

…index functionality

…umns