DeID improvements

[tomolopolis]

Small improvements for fine-tuning and using the de-id in a pipeline:

1. train / set splits can be performed outside of the train method. Changes to support this in DeIdModel.train
2. locale specific regex can be useful in a pipeline rather than directly collecting annotations and fine-tuning the underlying model. Changes to include arbitrary patterns matched and mapped to CDB cuis, then merged with model predictions. Eval code also doesn't use tokenizer split, so is fully representative of what was annotated.

[mart-r]

A few comments that I think should be addressed (i.e duplicate/typod key, remaining old doc string), the rest of it is more optional I'd say.
Though I do feel like the 2 comments regarding cui2preferred_name as well as the one regarding using filter make sense to implement as well.

[mart-r]

Some tests as well - great!
All good on my side.