Fix: Force Bfloat16 on CPU to satisfy PyTorch autocast requirements

Fix: Monkeypatch torch.cuda to prevent crash on non-NVIDIA CPU systems

Fix: Use git URL for hloc installation

Fix: Add missing hloc dependency for BA validation

Fix: Add robust device fallback (MPS -> CUDA/CPU)

Fix: Add missing h5py dependency

Fix API URLs: Use relative paths for deployment

Fix Dockerfile: Upgrade to Node 22 for Next.js 16

Fix Dockerfile: Allow yarn install without frozen lockfile

Fix Dockerfile: Replace libgl1-mesa-glx with libgl1

Clean deployment build (Squashed)