In-memory per-run LLM usage summaries.

Provider usage is also emitted to logs, but this GenServer keeps in-flight run-level cost and token accounting available to /api/runs. Completed run envelopes include the summary so persistence can retain the final accounting.