feat: add text input mode and fix TTS thread synchronization

[aopstudio]

Summary

• Add --text parameter for direct text input testing without requiring audio input
• Add eval_text_string helper function and test_case_text for text input mode
• Fix TTS/T2W thread synchronization issue: properly wait for threads to finish processing before sending stop signals
• Update README.md with --text option documentation and usage example

Changes

New Feature: Text Input Mode

Users can now test the model with text input directly, without needing to prepare audio files:

./build/bin/llama-omni-cli \
    -m /path/to/MiniCPM-o-4_5-gguf/MiniCPM-o-4_5-Q4_K_M.gguf \
    --text "Hello, please introduce yourself"

Bug Fix: TTS Thread Synchronization

Previously, omni_stop_threads was called immediately after test_case returned, which caused the TTS/T2W threads to be stopped before they finished processing. This resulted in incomplete or missing WAV output files.

The fix adds proper wait logic:

1. Wait for speek_done flag to indicate TTS completion
2. Wait for T2W queue to be empty before stopping threads
3. Add small delay for T2W to finish final processing

Test Plan

• Test text input mode with Chinese text: --text "请介绍一下Python语言"
• Verify WAV files are generated in tools/omni/output/round_000/tts_wav/
• Verify all threads terminate cleanly without hanging

🤖 Generated with Claude Code