[feature] decouple llm.resource from llm_knowledge

there are two use cases for llm.resource
- first basic case, a markdown representation of an oodo record (resource) extracted and formatted suitable for LLM context
- second, is to be used in semantic (rag) searches as a base for the chunks to point to

for the basic we may not even need a separate model - we can have a computed field md_content that with store=True for ir.attachment and store=False (dynamic) for all other models.

in case of ir.attachment we still want to be able to process them in batches using something like mistral_ocr but we do not necessary need to have llm_knowledge installed for this simple case

TBD