[Arjun99ab]

Connect the A4 audEERING tone model to the Mac server, so it returns real tone tags (excited, upset, calm, or none) instead of "no tone".

Changes

• tone.py: A4 tag thresholds; TONE_BACKEND picks mock (default) or audeering
• audeering.py (new): Runs the audEERING model; audio stays in memory
• main.py: Use the selected backend instead of always the stub
• test_tone.py (new): Threshold, backend, and end-to-end tests
• eval_tone.py, fetch_crema_d.py (new): Accuracy evaluation on labeled clips
• Docs: Updated thresholds and run instructions

Testing

• ruff and pytest pass
• P3 replay script gets real tags back
• ~45 ms per segment on Mac GPU
• CREMA-D (1,314 clips): 51% accuracy vs 33% always-none baseline; wrong tag on 16% of clips

When testing the main weakness was that happy speech is often tagged "upset". The thresholds need tuning on team recordings.