Bring back to_arrow_table and to_arrow_reader on DB-API results over a re-feed of the rows - #702

Brings back to_arrow_table(batch_size) and to_arrow_reader(batch_size) on DB-API results, on the connection and as module functions, over a re-feed of the rows the result holds.

The engine fixes a result's format at submission and execute() submits in the chunk format, so a result can no longer become an Arrow one. Instead, the rows the result still holds are moved into one ColumnDataCollection and submitted again in the Arrow format on the same connection, through the Session::Submit over a collection that 1dd1ab5 deleted. The conversion is the one the relational API's Arrow terminals use, so the two agree exactly: column names (duplicates included), batch sizes, types and settings. Each call emits a UserWarning pointing at con.sql(query).to_arrow_table(); deprecating or promoting them is for after 2.0.0.

Behavior against 1.5.x:

• to_arrow_reader no longer streams: the query runs to its end before the first batch. A table takes about 1.5x the relational call's peak memory, same time.
• After a partial row fetch, Arrow returns the remainder instead of raising "already fetched".
• With preserve_insertion_order = false the rows come back in a different block order per call.
• An aborted transaction, another open result and CONNECT mode refuse before anything is taken, rows still fetchable.