Chunked map-reduce for transcripts exceeding the context window

[wynnwu]

For very long transcripts: split -> summarize chunks -> reduce into one final summary, so inputs larger than the model context still work.