migrate to uv + pyproject.toml

- root pyproject.toml replaces requirements.txt and requirements-worker.txt
  (worker = root + ffmpeg-python which root already had); test deps moved
  to [dependency-groups] dev
- core/gpu/pyproject.toml replaces core/gpu/requirements.txt; uses
  [tool.uv.sources] to pin torch/torchvision and paddlepaddle-gpu to their
  CUDA index URLs, replacing the manual reinstall dance from old comments
- Dockerfiles use uv sync --frozen against uv.lock for reproducible builds;
  PATH includes /app/.venv/bin so k8s manifests' bare uvicorn/celery
  commands resolve without wrapping in uv run
- core/gpu/run.sh local mode now does uv sync + uv run python server.py;
  errors out cleanly if uv is missing

core/gpu/pyproject.toml

@@ -0,0 +1,42 @@
+[project]
+name = "mpr-gpu"
+version = "0.1.0"
+description = "MPR remote inference server (GPU)"
+requires-python = ">=3.11"
+dependencies = [
+    "fastapi>=0.109.0",
+    "uvicorn[standard]>=0.27.0",
+    "rapidfuzz>=3.0.0",
+    "Pillow>=10.0.0",
+    "redis>=5.0.0",
+    "ultralytics>=8.0.0",
+    "paddleocr>=3.0.0",
+    "paddlepaddle-gpu==3.0.0",
+    "transformers>=4.40.0,<5",
+    "accelerate>=0.27.0",
+    "torch",
+    "torchvision",
+    "opencv-python-headless>=4.8.0",
+]
+
+# RTX 3080 / CUDA toolkit 12.8 — cu126 wheels are forward-compatible
+# (no cu128 wheels yet on either index). Mixing PyPI torch with CUDA 12.8
+# causes NCCL symbol errors, so the explicit index pins prevent uv from
+# pulling torch transitively from PyPI via ultralytics.
+[tool.uv.sources]
+torch = { index = "pytorch-cu126" }
+torchvision = { index = "pytorch-cu126" }
+paddlepaddle-gpu = { index = "paddle-cu126" }
+
+[[tool.uv.index]]
+name = "pytorch-cu126"
+url = "https://download.pytorch.org/whl/cu126"
+explicit = true
+
+[[tool.uv.index]]
+name = "paddle-cu126"
+url = "https://www.paddlepaddle.org.cn/packages/stable/cu126/"
+explicit = true
+
+[tool.uv]
+package = false