Align on architecture for running LLM-as-judge evaluation

[aittalam]

See rationale here.

The main goal of this task is to end up with alignment on how we want to implement this new evaluation (e.g. as a new workflow which calls inference + LLM judge inference + evaluation vs having evaluation call LLM judge inference).
The deliverable should be a diagram explaining how we will run this new feature in Lumigator