Predicting medical student performance from attributes at entry: a latent class analysis