[REF] account: reconciliation models: improve performance

When opening the reconciliation widget and calling the matching rules, instead of calling all the rules in a single huge SQL query and then iterating on all the statement lines to match them with the results, we now first group the lines per applicable model and each time call a distinct (but way smaller) query for each group, stopping trying other rules for a statement line when one returns candidates. More queries are performed, but they are way more simple.