Add smoke tests for real API connectivity

Add a new smoke-api test suite that proves the CLI can complete real LLM
round-trips against the Anthropic API. Tests cover headless mode,
serve mode, and all-tools execution. Runs sequentially to avoid rate
limits, skips gracefully when API keys are missing (fork PRs), and
is wired into CI as a required check with allowed-skips.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>