Load cuda at runtime. Include cuda header files inside the project

TODO

@@ -1,6 +1,5 @@
 Check for reparent.
 Only add window to list if its the window is a topmost window.
-Load cuda at runtime with dlopen.
 Track window damages and only update then. That is better for output file size.
 Getting the texture of a window when using a compositor is an nvidia specific limitation. When gpu-screen-recorder supports other gpus then this can be ignored.
 Remove dependency on glfw (and glew?).