Reduce startup GPU quota burn: skip CUDA pkg compile if already importable

_install_cuda_packages now checks if nvdiffrast/diso/detectron2 are
importable before requesting a GPU. If packages survived the restart in
site-packages, skips the 120s GPU allocation entirely. Also reduces
@spaces.GPU duration from 300s to 120s (actual compile time is ~60-90s).

Each Space restart was burning 300s of ZeroGPU quota on recompilation
even when packages were already installed.

app.py

@@ -144,9 +144,21 @@ from PIL import Image
 # Compile nvdiffrast / detectron2 / diso here, on first GPU allocation at startup.
 _CUDA_PKG_MARKER = Path("/tmp/.cuda_pkgs_installed")
 
-@spaces.GPU(duration=300)
+def _cuda_pkgs_already_installed() -> bool:
+    """Return True if CUDA packages are importable (persisted in site-packages across restart)."""
+    try:
+        import nvdiffrast.torch  # noqa: F401
+        import diso              # noqa: F401
+        import detectron2        # noqa: F401
+        return True
+    except ImportError:
+        return False
+
+@spaces.GPU(duration=120)
 def _install_cuda_packages():
-    if _CUDA_PKG_MARKER.exists():
+    if _CUDA_PKG_MARKER.exists() or _cuda_pkgs_already_installed():
+        print("[startup] CUDA packages already installed — skipping compilation.")
+        _CUDA_PKG_MARKER.touch()
         return
     print("[startup] Installing CUDA packages (nvdiffrast, detectron2, diso)...")
     import shutil as _shutil