Error Bounds for Learning with Vector-Valued Random Features