During supervised training the document and entity provided to the trained component (linker only for now) don't align with each other.
There are 2 distinct issues:

• Since the document is generated from the regular inference procedure, it has .ner_ents and .linked_ents lists, but they are not aligned with what the annotated dataset specifies.
• The entity passed to the trained component is always a new instance with new state instead of being reused

This PR fixes the above by:

• Preparing the .ner_ents and .linked_ents for the document when doing supervised training (they will contain the same entities for now)
• Reworking how entity creation is done in this context in order to be able to reuse these entities (so the MutableEntity is within the .linked_ents)
  • This involved creating a new method entity_from_tokens_in_doc for the pipe and tokenizers
  • And deprecating the old one (entity_from_tokens)
• It also adds a few tests to support the above and updates other tests along with the changes