refactor: use dspy.context embedder

[enitrat]

Changes Made:

1. pgvector_rm.py:76 - Removed openai_client parameter entirely and made embedding_func optional, defaulting to dspy.settings.embedder
2. server/app.py:187 - Configure embedder once at startup in dspy.configure() alongside LM and adapter
3. server/app.py:649 - Removed embedder creation in lifespan(), now uses context embedder
4. document_retriever.py:549,619 - Removed hardcoded embedder instances, now relies on dspy.settings.embedder
5. All 3 optimizer files - Updated to configure embedder in dspy.configure() instead of passing directly to SourceFilteredPgVectorRM