Semi-supervised learning with pseudo-labeling compares favorably with large language models for regulatory sequence prediction