Language Confusion Gate: Language-Aware Decoding Through Model Self-Distillation