CC @asmorkalov @stopmosk

I confirm lazy loading with Ninja generator. It works with this PR.

But could you please provide steps to check python bindings with CUDA delayed loading?

Sure, no problem. If you compiled OpenCV with Python3 bindings enabled, go to your Python3 package location defined by PYTHON3_PACKAGES_PATH and open the cv2\python-X.X directory (depending on your Python version). Here, the cv2.cpXX-win_amd64.pyd file is a DLL with pyd extension whose dependencies can be checked just like those of other DLLs. In the existing code, the delay flags were only set for shared modules (i.e. DLLs) while with this PR, the delay flags are also set for modules (e.g. Python) as well as programs (i.e. EXE files).

Still, loading the pyd file from Python is a bit tricky regarding the dependencies. Even though all dependencies might be shipped next to the DLL or the Path environment, Python will not load these. Instead, they need to be defined by the package's configuration scripts (cv2\config.py here). But if you only check for delay-loaded dependencies, this should not really matter at all.