Recompute Interview Cost

Re-price every session in this interview using the current rates blob.

Motivating case: the agent stamps cost_breakdown onto each session at end-of-call using the rates active at that moment. If pricing_rates was updated later (admin PATCH, new models added to the seed), the stored totals are stale — typically $0 for any model that wasn't yet priced. The raw token counts, audio durations, and character counts ARE preserved per-turn, so we can recompute cleanly from stored data.

Idempotent. Running twice produces the same result. Does NOT recompute the interview's processing_cost (output generation) — that's a separate pipeline; we can add a sister endpoint if it becomes a real need.