This adds a possibility to take a connection from the pool to use it
via the Executor interface, and return it later when it's no longer
needed. This avoids connection pool overhead in cases when a lot of
quries need to be made, but the use of read transactions is not
needed.

Using read transactions instead of simple connections has the side
effect of blocking WAL checkpoints.

Using single connection for multiple SQL queries which are executed
during sync avoids noticeable overhead due to SQLite connection pool
delays.

Also, this change fixes memory overuse in DBSet. When initializing
DBSet from a database table, there's no need to use an FPTree with big
preallocated pool for the new entries that are added during recent sync.