Summarisation should use batch methods

[zacps]

Currently summarisation makes one call to the underlying LLM per document to summarise.

Instead aget_evidence should use batch methods (agenerate/aapply) to produce a single LLM call.

This could provide significant speedups for local LLMs.