Revisiting K-mer Profile for Effective and Scalable Genome Representation Learning