fix(memory): reduce psram pressure in chatbot runtime

[ronger-x]

Summary

• reduce display-side PSRAM pressure in the chatbot runtime by avoiding the full-height double-buffer profile on SPI panels
• keep audio helper worker stacks in internal memory and close the recorder with an internal stack when flash-unmapping paths are involved
• stop decoder resources promptly when returning to listening mode and lazily restart the decoder only when new downlink audio arrives
• move xiaozhi vision request body allocation to PSRAM when available to avoid consuming scarce internal RAM

Notes

• this PR focuses on runtime memory stability only
• board-specific SPIRAM default tuning for
  ymcu_bigsmart is tracked separately on codex/pr4-bigsmart-spiram

Validation

• python -m pre_commit run astyle_py --files agent/brookesia_agent_manager/src/base.cpp agent/brookesia_agent_xiaozhi/xiaozhi/src/esp_xiaozhi_vision.c examples/agent/chatbot/main/modules/display/display.cpp service/brookesia_service_audio/src/service_audio.cpp service/brookesia_service_helper/include/brookesia/service_helper/audio.hpp
• no build/test run in this pass