Depth Improves Training-Time Scaling for Difficult Tasks in Deep Linear Networks