Benchmarking Classical, Deep, and Generative Models for Human Activity Recognition