GPU-accelerated phash + fix discovery/takeout hang

GPU:
- Switch Dockerfile base to pytorch/pytorch:2.3.1-cuda12.1-cudnn8-runtime
- Add gpu_hasher.py: batched 2D DCT on GPU via PyTorch matrix multiply,
  256 images/batch, produces imagehash-compatible 64-bit hex hashes,
  auto-falls back to CPU when CUDA unavailable
- Replace per-image phash loop in scanner.py with phasher.hash_files()
- docker-compose.yml: add nvidia GPU device reservation

Hang fix:
- takeout.is_takeout_folder() now caps at 50 directories (was walking
  entire tree — blocked for minutes on 65k+ file libraries)
- Add "Not a Takeout folder" status message so takeout phase is never silent

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

requirements.txt

@@ -1,3 +1,7 @@
+# torch + torchvision come pre-installed in the pytorch/pytorch base image
+# (torchvision needed for image transforms)
+torchvision==0.18.1
+
 fastapi==0.115.6
 uvicorn==0.32.1
 Pillow==11.0.0
@@ -5,3 +9,4 @@ imagehash==4.3.1
 pillow-heif==0.21.0
 jinja2==3.1.4
 aiofiles==24.1.0
+numpy==1.26.4