Estimating Conditional Average Treatment Effects via Sufficient Representation Learning