BEND: Benchmarking DNA Language Models on biologically meaningful tasks