Plan a Map-Reduce Summarisation

To summarise a document far bigger than the context window: summarise groups of chunks (map), then summarise the summaries (reduce), repeating until one call can see everything.

Implement plan(chunk_tokens, context_limit, summary_tokens, overhead):

• Each call's input is its items' tokens plus overhead (instructions), and must fit in context_limit.
• Group items greedily in order: keep adding the next item while the call still fits. A single item that can't fit raises ValueError.
• Each call outputs one summary of summary_tokens tokens.
• Level 0 groups the chunks. Each later level groups the previous level's summaries. Stop after a level with a single call.
• If summary_tokens is so large that a level doesn't shrink the number of items, raise ValueError (it would never finish).

Return a list of levels, each a list of group sizes (items per call). Also implement total_calls(levels).