tflite interpreter for runtime inference

[cpmpercussion]

With the capacity to save models as tflite files we have experimented with making predictions using Tensorflow Lite.

On a powerful laptop, the present Keras model inference has an average time of 4.96ms (median 4.98) over 3200 predictions. Over 10000 predictions with a 9d s network, the tflite inference runner averages 0.17ms per prediction.

29.17x improvement!

Far too much performance improvement to leave on the table. This issue proposes building out the tflite inference engine so that it is automatically is used if a user specifies a .tflite file in config.toml.

If tflite inference seems to behave similarly, we should be doing this by default.

Thanks Scott H for helping to surface this.