Guard streaming prompts against context growth

[xiaowangzhixiao]

Summary

• add opt-in recent-turn and recent-image windows for main-model prompts
• preserve full chunk/session state for summarization and long-term memory
• roll back the current turn when prompt construction or the main-model request fails
• expose the new settings through CLI arguments, environment variables, and start_adapter.sh
• document the behavior in English and Chinese and add regression tests

Both limits default to 0 (unlimited), so existing deployments keep their current prompt behavior.

Fixes #35.

Validation

• python -m unittest discover -s services/webinfer/tests -v (5 tests passed)
• python -m py_compile services/webinfer/live_adapter.py services/webinfer/tests/test_live_adapter.py
• bash -n services/webinfer/scripts/start_adapter.sh
• git diff --check
• local end-to-end adapter run with 9 sequential image turns against an OpenAI-compatible vLLM backend; the final queried turn completed successfully
• forced backend failure followed by a retry stayed at the same turn/frame count, confirming rollback