feat(loops): surface tokenUsage on LoopResult + reportLoopUsage bridge for runProfileMatrix

[tangletools]

What

Makes runLoop surface the data agent-eval's new runProfileMatrix needs to run its backend-integrity guard — answering "how does agent-runtime consume the new primitive."

Why

runLoop already extracted per-call tokensIn/tokensOut from llm_call events, but only aggregated costUsd — the token counts were dropped before reaching Iteration or LoopResult. A product's runProfileMatrix/runCampaign dispatch that wraps runLoop could report cost but had no tokens to report, so agent-eval's assertRealBackend (which keys on tokenUsage) would misread a real run as a stub and throw.

Changes

• Iteration + LoopResult gain tokenUsage: { input, output }, summed per-iteration (across llm_call events) and across iterations (on the result). Structurally matches agent-eval's RunTokenUsage/CampaignTokenUsage.
• New reportLoopUsage(cost, result) — forwards a finished loop's cost and tokens into a campaign cost meter in one call. This is the trivial consumption path:

  const dispatch: ProfileDispatchFn<S, A> = async (profile, scenario, ctx) => {
    const result = await runLoop({ ...optsFor(profile, scenario), ctx: loopCtx })
    reportLoopUsage(ctx, result)   // cost + tokens → integrity guard sees real activity
    return result.winner?.output as A
  }

  Typed structurally (UsageSink) so loops/ stays free of an agent-eval import.

Verification

• typecheck + build green; full suite 381/381.
• Extended the existing cost-aggregation test to assert token aggregation (per-iteration + total) and reportLoopUsage forwarding — the regression being "a real run misread as a stub."

Note: agent-runtime does not import runProfileMatrix (products do); its role is to surface the data. The @tangle-network/agent-eval dep bump (^0.54.0 → latest) is a separate follow-up once runProfileMatrix is published.