Rework find_and_delete_entries (bugfix and better testing)

[hemidactylus]

This PR vastly improves the quality, reliability and behaviour of the find_and_delete_entries in the metadata mixin.

• The batch_size parameter was effectively ignored. FIX: it is no more
• The testing was scarce. FIX: now various call patterns are tested, with cross-checks by counting the items in the table before-and-after

But most importantly, a major bug has been fixed:
The find_and_delete_entries did not work when in combination with the clustered mixin and/or the multiple primary key mechanism. The reason was that the collection of primary keys (to be later used in deleting) did use the query results after the normalization, so that the flat structure of the table had been made into e.g. {"partition_id": ("x", 10), "row_id": ("a", 1), ...} in case of multicolumn (primary/clustering) key. The deletions then failed miserably.

• Now the find_entries is split into a pre- and a post-row-normalization parts, so that the find_and_delete_entries can use the former and get access to the actual columns, which is what it needs for the later delete statements.
• Also comprehensive tests for this column schema have been added.

As a final note, everything in this PR is doubled in its asyncio counterpart.