docs: add provider-neutral AgentsKit eval cookbook

[Emerson Braun (EmersonBraun)]

Summary

• add a deterministic TypeScript support agent and three evaluation fixtures
• run the same baseline locally with @agentskit/eval and through Braintrust
• add required-term and safe-escalation scorers with credential-free tests
• register the cookbook and Emerson Braun as the author

Why

This gives Braintrust users a provider-neutral evaluation recipe that can run before
they have a model key or production traces. The async task boundary can later be
replaced by any AgentsKit runtime, hosted service, or model provider without rewriting
the dataset or scorers.

Validation

• pnpm install --frozen-lockfile
• pnpm --filter agentskit-braintrust-eval test — 3 passed
• pnpm --filter agentskit-braintrust-eval check
• pnpm --filter agentskit-braintrust-eval eval:local — 3/3 passed, accuracy 1
• braintrust eval agent.eval.ts --list --no-send-logs
• Prettier and registry/author schema-shape validation passed