Li fix

[liyin2015]

Summary by CodeRabbit

• New Features
  • Introduced flexible, turn-based conversation memory with per-turn grouping, message formatting, and retrieval.
  • Improved streaming stability: better final-answer detection, reliable “execution complete” event, and smoother event-loop pacing.
  • Enhanced structured-output generation for more consistent JSON responses.
• Bug Fixes
  • Adding an assistant response without a pending user query now logs a warning instead of raising an error.
• Documentation
  • Added examples and scripts to compare structured-output approaches and an experimental data-extraction workflow with a sample dataset.
• Tests
  • Extensive new tests covering streaming (sync/async), multimodal inputs, reasoning models, headers/organization, and runner finalization.