[query] Revisit emitting ApplyIR nodes as separate methods

In #8067, we made a change to always inline ApplyIR. This was done for technical reasons, but as noted there, we really should be emitting ApplyIR nodes as separate methods. In particular, this should help us avoid class-too-large exceptions. We should investigate whether those technical barriers are still there, and how much work it would now take to achieve the desired behavior.