Policy Gradient-Based EMT-in-the-Loop Learning to Mitigate Sub-Synchronous Control Interactions