Sample-efficient inverse design of freeform nanophotonic devices with physics-informed reinforcement learning | Synapse