vclmax/Confucius4-TTS

Grow DiT rotary cache for long mel sequences; raise GPU budget to 60s

Use ZeroGPU duration=30s (x1.5 factor = 45s, under the 60s free cap)

Load model on CPU at startup; always re-assert GPU placement

Add vendored torchaudio shim for ZeroGPU runtime

Confucius4-TTS: English UI, arbitrary audio upload, ZeroGPU