feat: prepare_table() makes O(1) calls to prepare_column(), instead of O(n) (n number of columns)

[raulbonet]

Feature scope

Taps (catalog, state, stream maps, tests, etc.)

Description

Function prepare_table() in SQLConnector calls prepare_column(), for each column in a table.
In turn, this function queries the database to first determine whether the column exists or not. Therefore we make O(n) calls to the database.
However, in a single call we can retrieve all columns and leverage this information.

The target-postgres already implemented this optimisation by overriding the default methods prepare_table() and prepare_columns().

The way target-postgres implemented it implied modifying the signature of the functions, my implementation is different.
I take the responsibility of checking whether the column exists out of the prepare_column() function and encapsulate this in a new function called prepare_table_columns().

However, a thought: I think that in the end the underlying issue is that functions like prepare_table() and this new function prepare_table_columns() belong to the SQLSink class rather than the connector: a new sink is created every time there is a schema change. The class that knows the current schema and therefore can do these optimisations is the sink, not the connector maybe?