Small Apache-2.0 Models Against Gemma 4 E2B on One TPU v5e: Which Ones Keep Up?