As long as the audio interface is fast enough there's no need to listen to the audio output of the TD-12 brain.
There are plenty of tips to reduce latency in the Sonar section of the forum. A few highlights being 1) to ensure you do not have any network adapter active (or wireless adapter) while recording/tracking and 2) to not have effects active on any tracks while recording.
However, that said, if the audio interface is not that snappy, then the converse strategy is to listen only to the TD-12 audio output (for the drums mixed in with the other playback tracks of the song), and not listen at all to AD2. That also is a reliable strategy. Then after recording the drums, you go back and listen to what you performed through AD2, and have the TD-12 output complete off. The disadvantage of that is you can't take advantage of what you might have set up in AD2 to hear while playing/performing the drums. The advantage is that you can be spot on with timing.
But, if you have next to no latency, as outlined at the start, it's better to only monitor through AD2 and have the TD-12 audio output completely silent while tracking/recording.