api-rs: persist usage_records for workflow agent turns

[ordemri]

Summary

Workflow-tier agent turns (ctx.agent_turn in the python workflow host) wrote no usage
observation at all — unlike session-tier turns, whose harness turn.completed events
already carry a usage object. Workflow turns already receive the same output-line
stream session turns do, so this extracts usage from that existing stream rather than
adding a separate accounting mechanism, and persists it as a new usage_records row per
completed turn: {execution_id, thread_key, harness, model, input_tokens, output_tokens}.

Gated behind the existing telemetry trace-export config (a no-op when trace export is
off, so it doesn't become a second, independent form of telemetry). Best-effort: a write
failure is logged, never propagated, so usage accounting can't fail a workflow's actual
agent turn.

Fixes #1111.

Changes

• New migration: usage_records table + index on execution_id.
• SessionRuntime::pool(): exposes the Postgres pool already backing the runtime's
  session store, so callers that share that store (like workflow-tier turns) can persist
  their own rows without a second connection pool.
• centaur-workflows: parses usage off turn.completed/turn/completed events in the
  turn's output-line stream and writes one row per turn.

Test plan

• cargo fmt --all --check
• cargo clippy --workspace --all-targets -- -D warnings
• cargo test --workspace (ParadeDB Postgres via start-paradedb-postgres.sh)
• New test: a workflow agent turn produces exactly one usage_records row with the
  expected token counts, model, and harness.