docs: add AGENTS.md with T4/headless-setup notes

[moduvoice]

Motivation

While benchmarking the AICoverGen Mod inference flow on a real NVIDIA Tesla T4 (Turing, 16GB), I hit several non-obvious setup issues that aren't documented anywhere in this repo. Adding an AGENTS.md so the next person running this outside Colab (or an AI coding agent doing the same) doesn't have to rediscover these from scratch.

Changes

Doc-only — new AGENTS.md, no code touched. It documents four things about existing behavior:

1. The documented flow is Colab-only. Hina_Mod_AICoverGen_colab.ipynb mounts Drive, hardcodes /content/... paths, and clones the actual inference code from a rot13-obfuscated URL (ardha27/AICoverGen-Mod). None of this runs as-is on a non-Colab GPU box. Notes the headless path (python src/main.py) and the dependency reconstruction needed (torch 2.0.1+cu118, python3.10-dev for building fairseq==0.12.2 from source, huggingface_hub==0.23.4, setuptools<81, ffmpeg/sox).
2. No bundled RVC voice model by default. download_models.py only fetches the MDX/hubert/rmvpe models, not an RVC voice model, so no_ui.py's default -dir Daemi points at a folder that doesn't exist and the run crashes immediately. Notes that users must supply their own RVC model.
3. MDX vocal separation silently falls back to CPU without extra CUDA runtime libraries (nvidia-curand/cufft/nvrtc, not bundled with the torch cu118 wheel). Measured 2.05x speedup (28.7s → 14.0s per separation pass) once supplied, with a VRAM trade-off (~2GB → ~13.8GB peak) worth flagging for 16GB cards.
4. fp16 is correct on T4 (no bf16 trap). Config hardcodes fp16, and the GTX 16xx fp32-downgrade heuristic checks for the substring "16" in the GPU name — "Tesla T4" doesn't match it, so T4 correctly stays fp16. Also notes that the 16xx downgrade branch, when it does trigger, rewrites config JSON files on disk as a side effect.

Testing

• Verified end-to-end on a real Tesla T4 (16GB, driver 550.163.01, torch 2.0.1+cu118): reconstructed the headless flow, generated a cover with a supplied RVC voice model, and measured MDX separation timing/VRAM under both the CPU-fallback and CUDA-library-supplied conditions.
• No code paths were touched; this is a documentation-only change and doesn't affect existing behavior.