More clear labelling of chunks and paragraphs when providing as evidence to LLM

[jrochkind]

To try to encourage more accurate and specific referents and citations.

Every chunk will now have an (eg) "ORAL HISTORY ID: OH0627" metadata line.

Every paragraph will begin with "[OH0627|P32] SPEAKER:" -- additional labelling of OH id and paragraph number on every paragraph; and every paragraph has a speaker label if we can provide one, not just assumed from prior paragraph.

• move chunk formatting for LLM into ERB template
• add OH numbers to chunks metadata provided to LLM
• in chunks to LLM add OH and Paragraph number before every paragraph, to try to improve correctness and specificity of citations
• When creating Chunk records, make sure every paragraph in text begins with a speaker label, assuming from previous paragraph if needed

---

After merge, need to re-create all chunks on staging (which will cost us maybe $20), to have properly labelled speaker paragraphs stored in chunk records

Bah, nevermind -- right now, deleting/regenerating chunks will cause all bookmarked questions to disappear, we don't want to do that without thinking and ideally a solution.

So this PR will include new chunk generation code, to ensure all paragraphs have a speaker label, but the chunks on staging won't necessarily reflect it yet. Any new chunks we make in future will. Other improvements in this PR will be immediate.

[jrochkind]

After merge, need to re-create all chunks on staging (which will cost us maybe $20), to have properly labelled speaker paragraphs stored in chunk records

Bah, nevermind -- right now, deleting/regenerating chunks will cause all bookmarked questions to disappear, we don't want to do that without thinking and ideally a solution.

So this PR will include new chunk generation code, to ensure all paragraphs have a speaker label, but the chunks on staging won't necessarily reflect it yet. Any new chunks we make in future will. Other improvements in this PR will be immediate.