Changes necessary for the Ensenso grabber

This MR changes the following:

* Add option to build OpenCV with CUDA support, and patch openCV to fix building with CUDA with jom and Ninja on Windows. This is caused by the default install dir for CUDA containing spaces... The patch has been [submitted upstream and merged](opencv/opencv#8819) 🎆 
* Patch PCL to find the Ensenso SDK on Windows because... Yeah, spaces. The patch has been [submitted upstream and merged](PointCloudLibrary/pcl#1875). 😄 
* Bugfix: in the PCL builder CMake code, replace some `BUILD_X` variables with no effect by the actual variables, `WITH_X` (eg. `BUILD_OPENNI` does nothing, the variable is `WITH_OPENNI`).
* Add option to build PCL without CUDA support, to be coherent with what this MR does for OpenCV (currently, PCL is built with CUDA support whenever CUDA is detected).
* Fix PCL and OpenCV not being able to build CUDA targets on windows in parallel. The speed gain here is pretty good  
  Before/After:
    * PCL: 21min/10min
    * OpenCV: 40min/12min
* Ability to set for which CUDA architectures to build dependencies using CUDA. This is not _necessary_, but the build time with CUDA is horribly long, even more so with OpenCV and CUDA support, and this speeds things up quite a bit. By default, build for all SM >= 3.0 architectures (2.x is old and deprecated, I'd be surprised if anyone used this). To speed things up, set it to the architecture of your card **(AND THE GPUS OF THE MACHINES YOU WILL PACKAGE INSTALLERS FOR)**  
  Building PCL + OpenCV + LibSGM (including the aforementioned parallelism improvement):
    * SM 6.0: 21min
    * SM 3.0 3.5 3.7 5.0 5.2 5.3 6.0 6.2: ~1h
* Add a new optional dependency, libSGM, and a patch that basically rewrites the build system from scratch. Also includes a cmake `FindLibSGM` module.
* Windows people who want to show appreciation for the CUDA build speed improvements may send pains au chocolat to @rchavignat. (a man can dream...)


See merge request !58