1. Support document model in llm_filter and some performance improvements
llm_filter will work on a DocSet made of original Documents (i.e. pre-exploded state) when using use_elements=True

• Element iteration "early stopping" - a record (document) passes the filter if any of the document's elements pass the llm_filter condition, we don't evaluate every element. This helps for documents which would be a hit, but doesn't for documents which would not (because we would still iterate over every element)
• Optionally, it sorts elements by a similarity score to get potential llm hits sooner
• keep_none flag to dictate behavior for a missing property value

None of these are perfect but they align with the use cases we have so far and can keep iterating.

2. Sorted elements returned by OpenSearchReader in reconstructed documents
Use element.element_index field to sort elements when reconstructing a document. Does an explicit sort after each reconstruction which is okay based on the magnitude of data.

Note: the ignore_doc_structure flag is to prevent current usage from breaking, once we update other code to use the document model we can remove/flip it