Detect CUDA home dynamically instead of hardcoded /cuda-image/... path

The hardcoded /cuda-image/usr/local/cuda-12.9 path does not exist on
the current ZeroGPU runtime container. Detect via shutil.which("nvcc")
first, then scan common /usr/local/cuda* paths as fallback.

app.py

@@ -128,8 +128,31 @@ def _install_runtime_packages():
     )
     # A10G = sm_86. Set arch list explicitly because NVML is unavailable at install
     # time (no GPU allocated yet), so torch can't auto-detect it.
-    # CUDA headers live at a non-standard path on ZeroGPU runtime containers.
-    _cuda_home = "/cuda-image/usr/local/cuda-12.9"
+    # Detect CUDA home dynamically — path varies between ZeroGPU container builds.
+    import shutil as _shutil
+    _nvcc = _shutil.which("nvcc")
+    if _nvcc:
+        _cuda_home = str(Path(_nvcc).parent.parent)
+    else:
+        # Scan common ZeroGPU / CUDA container paths
+        for _cand in [
+            "/usr/local/cuda",
+            "/usr/local/cuda-12.9",
+            "/usr/local/cuda-12.8",
+            "/usr/local/cuda-12.4",
+            "/usr/local/cuda-12.1",
+            "/cuda-image/usr/local/cuda-12.9",
+            "/cuda-image/usr/local/cuda",
+        ]:
+            if Path(_cand, "bin", "nvcc").exists():
+                _cuda_home = _cand
+                break
+        else:
+            raise RuntimeError(
+                "nvcc not found — no CUDA toolkit in PATH or standard locations. "
+                "Cannot build nvdiffrast/detectron2/diso."
+            )
+    print(f"[startup] CUDA home: {_cuda_home}")
     _cuda_env = {
         **os.environ,
         "TORCH_CUDA_ARCH_LIST": "8.6",