FIX: Ensure conversation scoring only adds once to memory

[jbolor21]

Description

Conversation Scorer adds to memory twice!

Previously, the ._score_async called the wrapped scorer's public score_async function, which persists scores to memory. This meant the outer Scorer.score_async (that invoked _score_async) then persisted them a second time, producing two identical ScoreEntry rows per call. The previous workaround regenerated each score's id via uuid.uuid4() after the inner call to dodge the collision, but this still meant the scores were still being written twice (with different IDs).

Fix:
Call the wrapped scorer's protected _score_async instead. This skips the inner persist (and the inner validator/role-filter/blocked-content substitution, which is safe here because the synthetic conversation message we build is pure text with response_error="none"). The outer Scorer.score_async now persists the returned scores exactly once, keyed to the original message_piece_id. The ID-regeneration workaround is removed.

Tests and Documentation

• pre-commit passes all unit tests
• reran notebooks with manual conversations
• New unit test confirms exactly one ScoreEntry is persisted per call with the original score ID intact