Inference for Heterogeneous Treatment Effects with Efficient Instruments and Machine Learning