Local Metric Learning for Off-Policy Evaluation in Contextual Bandits with Continuous Actions