python3Packages.onnxruntime: buildInputs dep on onnxruntime

The onnxruntime libs are patched to look for additional runtime
libraries in `${onnxruntime}/lib`, so we must depend on it.

Alternatively, onnxruntime could be built twice, once for each lib path
patch, but onnxruntime takes a vert long time to build with cuda support.

pkgs/development/python-modules/onnxruntime/default.nix

@@ -50,6 +50,13 @@ buildPythonPackage {
       oneDNN
       re2
       onnxruntime.protobuf
+
+      # https://github.com/NixOS/nixpkgs/pull/357656 patches the onnx lib to ${pkgs.onnxruntime}/lib
+      # but these files are copied into this package too. If the origional non-python onnxruntime
+      # package is GC-ed, cuda support in this python package will break.
+      # Two options, rebuild onnxruntime twice with the different paths hard-coded, or just hold a runtime
+      # dependency between the two. Option 2, because onnxruntime takes forever to build with cuda support.
+      onnxruntime
     ]
     ++ lib.optionals onnxruntime.passthru.cudaSupport (
       with onnxruntime.passthru.cudaPackages;