Learning to Recover from Plan Execution Errors during Robot Manipulation: A Neuro-symbolic Approach