backlog: follow up on search provider config

1. It feels a bit wrong for an object to mutate its data object passed in during initialization. I'd prefer to handle this at the Pydantic validation stage.
2. Some embedding models (Gemini) have a dimensionality value in embedding_parameters which we'd need to add to item hash. Without this if a users changes the dimensionality we won't invalidate the cache and cause errors.

Originally posted by @tgasser-nv in https://github.com/NVIDIA-NeMo/Guardrails/pull/1753#pullrequestreview-4052578064

Review PR #1753 and trace the search provider configuration and Pydantic validation paths. Check how embedding_parameters contribute to the item hash and cache invalidation, including the dimensionality value described in the issue. Done means initialization no longer mutates the passed data object and dimensionality changes invalidate cached items.