[Caspar] Replace legacy FindCUDA with modern CMake CUDA language support + minor compile function improvements

[tordnat]

• Replaces deprecated FindCUDA with CMake's native CUDA language support (requires CMake 3.18)
• Fixes non-portable behavior from -arch=sm_75. It embeds only sm_75 SASS with no PTX, so on newer architectures the driver silently falls back to running Turing SASS under backward compatibility. I experienced that results can differ across GPU generations when this flag was set. The new CMAKE_CUDA_ARCHITECTURES embeds SASS for sm_75/80/86/89 and PTX for forward JIT-compatibility on future architectures. The logic is to by default have maximum compatability, and let the user specify the specific arch if they want to reduce binary size etc.
• Respects CUDACXX env var before project() to handle systems with conflicting CUDA installations
• CasparLibrary.compile() gains cuda_architectures and jobs parameters. The latter defaults to a single job to avoid OOM on machines where parallel CUDA compilation exhausts RAM (I only have 16gig rn, so this was an issue even with just a small amount of kernels)
• Fixes CMake dev warning by adding DOWNLOAD_EXTRACT_TIMESTAMP TRUE to the pybind11 FetchContent_Declare
• Some minor improvements to the compile() function hopefully making it easier to build

I have tested this on sm_86 and sm_89 GPUs.

[tordnat]

For reference: This was discovered while unit testing the ongoing integration of Caspar into Colmap. Would be very nice to have build files that work more nicely with larger project, hence my contribution. colmap/colmap#4018

[aaron-skydio]

Nice, just one question, otherwise looks good - thanks for the excellent description

[aaron-skydio]

Thanks, I'll get this merged

[tordnat]

Sorry for the last minute change. Added an option to make compiling the Python bindings optional, as some projects may only use the C++ library.

[tordnat]

Sorry for the last minute change. Added an option to make compiling the Python bindings optional, as some projects may only use the C++ library.

This should also be an option in the generate function. Will make a new PR based on this branch.