fix: Improve audio splitting robustness and encoding handling / fix(tts): Handle reference audio prerequisites for GPT-SoVITS batch processing

[mthezi]

• Add explicit UTF-8 encoding with fallback for FFmpeg output
• Implement comprehensive error handling for silence detection
• Add robust parsing of FFmpeg silence detection output
• Fallback to fixed-length splitting when silence detection fails
• Add detailed warning messages for debugging purposes

This commit fixes the GBK codec error and makes the audio splitting process more reliable by gracefully handling potential failures in the FFmpeg silence detection process.

[mthezi]

• Add reference audio existence check for REFER_MODE 2/3
• Add automatic extraction when reference audio is missing
• Implement REFER_MODE 3 fallback to REFER_MODE 2 on failure
• Ensure refers are generated before batch TTS processing

The secondly fix.