Add HolmesVAU Gradio Demo App + Backend Improvements (GPU, Indexing)

[jungseoik]

✨ Summary

This PR introduces a Gradio-based demo interface for HolmesVAU and includes multiple improvements to enhance backend compatibility, robustness, and usability.

---

Commits Overview

✅ [fixed] Make anomaly scorer GPU allocation dynamic instead of fixed to GPU 0 and 1

• Refactored torch.load(...) with map_location=self.device to ensure the anomaly scorer loads properly on any available GPU setup.
• Previously hardcoded to cuda:0,1, which was incompatible with some environments.

✅ [fixed] frame and patch list indexing by casting sampled indices to int

• Fixed a subtle bug where sampled_idxs may contain float values, which are unsafe for direct indexing.
• Now explicitly casts each index to int for stable frame/patch retrieval.

✅ [fixed] pydantic==2.10.6 dependency add

• Added a fixed version of pydantic to ensure consistent behavior with Gradio & FastAPI.

✅ [feat] gradio demo app add

• Introduced app.py with a full Gradio Blocks interface:
  • Upload video input
  • Enter custom prompt
  • Get model prediction
  • See anomaly plot or "uniform sampling" flat plot
  • Browse keyframes via slider

[add] demo mp4 add

• Included a sample .mp4 video for testing/demo purposes.

---

🖥️ Demo Features

• 🎥 Video Upload + Prompt Input
• 📈 Anomaly Score Plot (flat if uniform sampling)
• 📸 Keyframe browsing with slider
• ✅ Supports cases with or without anomaly scores

---

📝 Notes

• Compatible with both single and multi-GPU machines
• Works with uniform or ATS-based frame sampling
• Keyframe images are saved to keyframes/ for interactive exploration

---

📸

holmes_vau_demo.mp4

---