docs: add Tesla T4 (Turing) hardware notes

[moduvoice]

Motivation

While verifying this project's inference pipeline on a real NVIDIA Tesla T4 (16GB, Turing/sm_75 —
a common budget/free-tier GPU, e.g. Google Colab's free tier), I found some hardware-specific
behavior and a couple of install-doc gaps that aren't currently documented. This PR adds that
documentation only — no code or behavior changes.

Changes

• Fix Manual Install Option 2 and Option 2b pip commands to include s3tokenizer==0.3.0
  and onnx==1.16.0 plus the protobuf upgrade — the same command start.py's automatic installer
  already runs (install_chatterbox_no_deps()), but the manual doc block was missing these two
  packages, so following it verbatim fails at import time with
  ModuleNotFoundError: No module named 's3tokenizer'.
• Add predefined_voice_id to the /tts streaming curl example (plus a short note on why it's
  required), since the example as written returns HTTP 400 {"detail":"Missing 'predefined_voice_id' for 'predefined' voice mode."} — voice_mode defaults to "predefined" and the server requires
  the id in that mode.
• Add a new AGENTS.md documenting that TTS_BF16 should stay off (the default) on T4/Turing
  (sm_75) GPUs: torch.cuda.is_bf16_supported() reports True on T4, so TTS_BF16=auto silently
  enables bf16, but Turing has no bf16 tensor cores and it measures ~1.6x slower than the fp32
  default. Also notes that the default Turbo engine ignores exaggeration/cfg_weight even though
  the API/config expose them as tunable.

Testing

• Measured on a real Tesla T4 16GB (Turing, sm_75), driver 550.163.01, torch 2.5.1+cu121, default
  Chatterbox-Turbo engine, commit 915ae28.
• Reproduced the manual-install failure verbatim: pip install -r requirements-nvidia.txt && pip install --no-deps git+https://github.com/devnen/chatterbox-v2.git@master → base ChatterboxTTS
  import fails with ModuleNotFoundError: No module named 's3tokenizer'. Adding
  s3tokenizer==0.3.0 onnx==1.16.0 + pip install --no-deps --force-reinstall "protobuf>=4.25.0"
  resolves it (ChatterboxTTS OK / Turbo OK).
• Reproduced the curl example verbatim → HTTP 400; adding "predefined_voice_id":"Emily.wav" →
  HTTP 200 (159KB wav, custom text/seed applied correctly).
• bf16 vs fp32 on T4, measured via engine.synthesize() directly (seed=1234, 5 warm runs each,
  <0.5% run-to-run variance): warm RTF 0.335 (fp32/default) vs 0.535 (TTS_BF16=on/auto) —
  bf16 is ~1.6x slower wall-clock per second of audio. Confirmed the bf16 path was actually active
  at runtime (t3 dtype = torch.bfloat16, torch.is_autocast_enabled()=True) rather than a
  no-op. Peak VRAM: 4.91GB (fp32) vs 3.92GB (bf16) — well within the 16GB budget either way, so the
  VRAM savings aren't needed on this card.