What does effective dimension reveal about redundancy, generalization, and trainability that parameter count, loss, and accuracy do not?