Universal Inference Meets Random Projections: A Scalable Test for Log-concavity