Matching on What Matters: A Pseudo-Metric Learning Approach to Matching Estimation in High Dimensions