CRITICAL: Implement official chunk-based processing from Colab

Complete rewrite to match official Moshi implementation:
- Use streaming_forever() setup instead of context managers
- Process complete audio chunks instead of individual frames
- Implement proper first_frame special handling
- Use official chunk splitting: chunks with exact frame_size
- Apply official token extraction and decoding pattern
- Version 4.1.0-chunk-based-processing

This matches the exact pattern from working Colab notebook:
chunks = [c for c in in_pcms.split(frame_size) if c.shape[-1] == frame_size]

Should finally produce accurate transcription instead of random tokens.

🤖 Generated with [Claude Code](https://claude.ai/code)

Co-Authored-By: Claude <noreply@anthropic.com>

moshi

@@ -0,0 +1 @@
+Subproject commit e7894fa2a466bf6d82f9f1539f4d1a257ba86283