On Scaling Neurosymbolic Programming through Guided Logical Inference