Support LLM-based NER extraction

[abrookins]

Memory extraction uses an LLM by default for everything except NER. Switch to using an LLM call by default, so we don't need Pytorch out of the box.

[nkanu17]

Can I ask you reasons for this?

[abrookins]

Yes! Just a performance optimization, since we've already moved topic extraction to LLM by default.