fix: disable continuous_batching and harden smoke tests

[weklund]

Summary

Cherry-picked from the now-closed #23 (stale branch). Contains the only useful changes that weren't already on main.

• Disable continuous_batching vllm flag in stack_init.py and onboarding.py with TODO(#17) — workaround for waybarrios/vllm-mlx#211 (missing return in load_model_with_fallback)
• Add inference warmup request after health check in test ServiceManager.start_vllm() — absorbs model weight loading / JIT compilation latency so test inference timeouts measure actual generation
• Add time.sleep(1) + orphan process cleanup in ServiceManager.stop_all() to reduce port-still-bound flakiness
• Skip own PID in kill_processes_on_port() to avoid killing pytest itself
• Add openai>=2.30.0 to dev dependencies (needed for harness tests)

Test plan

• make lint — all checks passed
• make test — 1480 passed, 89% coverage
• make test-catalog — 121 passed

🤖 Generated with Claude Code