Need to change the spaCy cli train code to save the each iteration training result to disk for NER training

[YanLiang1102]

In that way we can see how each iteration enhance the tag and token accuracy since sometimes I can see that the later iteration actually get a worse result...

[ahalterman]

Good idea! Or you can do it Keras-style: generate the validation loss/accuracy and only save the model if it outperforms the best previous iteration.