feat: add robust logging and diagnostics for ZeroGPU troubleshooting; fix '0 seconds' issue with explicit checks

fix: move all output and storage directories from /tmp to local app subdirectories to fix Gradio serving issues on HF; explicitly allow results/checkpoints paths; set Audio type to filepath

fix: lazy-import torch/torchaudio in inference.py to avoid libcudart.so.13 crash; pin cu124 wheels in requirements to prevent demucs pulling CUDA-13 torchaudio

Add voice similarity control + improve reference audio processing

Upgrade audio quality: pro mixing chain, better inference params, htdemucs_ft

Fix GPU timeout: increase to 300s, lower default diffusion steps to 10

Fix inference tensor backward error: add torch.no_grad + inference_mode

Add GPU worker debug logging to diagnose ZeroGPU import errors

Fix missing 'munch' dep + add detailed error wrapper for ZeroGPU

Bundle Seed-VC modules directly in repo for ZeroGPU compatibility

Fix ModuleNotFoundError: clone Seed-VC to app dir instead of /tmp

Replace RVC with Seed-VC for zero-shot voice conversion

Implement real RVC v2 inference pipeline with HuBERT + FAISS + generator

Convert pretrained checkpoint to RVC inference format for proper voice conversion

Upgrade to Gradio 5.12.0 for proper ZeroGPU authentication

Remove @spaces.GPU from separation and inference for CPU fallback

Rewrite inference with fallback: try Applio, then pitch-shift + FAISS

Reduce all @spaces.GPU durations to 60s (ZeroGPU max limit)

Initial commit: Clone Vocal RVC - web voice cloning tool