Towards Optimal Adapter Placement for Efficient Transfer Learning