Ship the modelling code, tokenizer and auto_map so the model loads with transformers alone

Publish the export rather than the training checkpoint