Refactor ReaderModificationCommandBatch

[roji]

As discussed offline with @AndriySvyryd, ReaderModificationCommandBatch currently adds a parameter for each IColumnModification, for each modification command - in order. This makes it difficult to implement patterns which do fancy things with parameters. For example, on PostgreSQL insertion of multiple rows can be optimized as follows:

INSERT INTO data_table (id, name) SELECT * FROM unnest(@i, @n)

Where @i and @n are array for the column values of the different rows.

In addition, I think we can simplify the current design and apply commands to the batch directly as they're added, as opposed to the current CachedCommandText mechanism.