[Feature]: Per-slot reasoning_effort for MoA reference advisors

MoA reference models spend most of their latency on private reasoning the aggregator never sees, with no way to reduce it independently; the proposal adds a per-slot reasoning_effort option, already implemented per PR #57043.