Context Search With Documents (Personalization) - Revised

[vicilliar]

Updated version of personalization PR intended for 2.21.0 which was reverted due to an issue with the Vespa client.
Differences from old personalization PR:
-Use persistent SSL context instead of persistent transport, because it caused hanging during concurrent requests. Performance gains are comparable.
-Integration test added for concurrent get documents requests to ensure no hanging

Change Summary

Supports new key documents in context parameter of search. This allows for a dict of doc_id:weight pairs to be provided by user for vectors to be extracted and interpolated for tensor search or hybrid search with tensor retrieval or ranking.

Updated example context object:

    {
        "tensor": [{"vector": [1 for i in range(768)], "weight": 1}], 
        "documents": {
            "ids": {"doc1": 1},
            "parameters": {
                "excludeInputDocuments": False,
                "tensorFields": ["text_field_1"],
                "concurrency": 10
            }
        }
    }

-Interpolation is now done on all vectors (query, context.tensor, context.document) instead of previous behavior of weighted average.
-Absolute value is now used when summing vector weights in interpolation (This makes ZeroSumWeightsError obsolete, we now use AllZeroWeightsError instead)
-Interpolation logic implemented in numpy.
-Fixed some error messages for hybrid / pydantic model validation.

New Env Vars:
-MARQO_MAX_SEARCH_CONTEXT_DOCS - aximum number of documents that can be used as context documents in a search request.

The following optimizations have been implemented:
-Getting doc vectors fetches only essential embeddings fields from Vespa
-A persistent SSL Context is created upon vespa client initialization and is used for the verify param of every async client instantiation. This saves time spent creating new contexts.
-Index is fetched from cache instead of Vespa
-Fetched doc vectors are not formatted into document form with Pydantic. Vectors are just used directly

Related Jira Ticket

https://s2search.atlassian.net/jira/polaris/projects/MOSD/ideas/view/4863474?selectedIssue=MOSD-23&issueViewLayout=sidebar&issueViewSection=overview&atlOrigin=eyJpIjoiNzcyNDBkNzVjM2JhNDE0Y2I4ODUzNDY3YTc2MmE0ZmUiLCJwIjoiaiJ9

Checklist

• Tests have been added for changes
• Documentation has been updated
• Breaking changes are clearly identified
• Python client changes linked or N/A
  py-marqo changes: Support for context documents in search (Personalization) py-marqo#287

For new field types:

• Tests cover score modifier usage of this new type
• Test indexes updated to cover the new type for all APIs (add docs, search, partial update, etc.)