Save memory by changing to float16

[florian-huber]

Currently, a large array of reference scores (e.g. Tanimoto) is used during training.
The data format can be downscaled to float16 to save memory (now: float32 or float64?).