Cycles: Improve sampling of area lights

This patch uses the sampling method described in "A Low Distortion Map Between Triangle and Square" by Eric Heitz.
The benefit is avoiding sqrt in the calculation, which could be cheaper on some architectures, and the result is
more even sampling across the triangle surface.

Based on ideas from
https://pharr.org/matt/blog/2019/02/27/triangle-sampling-1.html
https://pharr.org/matt/blog/2019/03/13/triangle-sampling-1.5.html

Reviewed By: Brecht Van Lommel

Differential Revision: https://developer.blender.org/D6566

Revert "Win32: Use the nicer looking blender-drawn confirmation message box when"

This reverts commit 60d6eb0b5d578eb02c5f1f31cb51a6cbeea6ee6f.

There is already a full patch (D3118) by brecht for this, will leave it up to him.

Win32: Use the nicer looking blender-drawn confirmation message box when
quitting with unsaved changes, instead of an ugly messagebox

Fix missing icons in non-portable local builds.

The problem is that some local path is always returned by
bpy_resource_path. The function does not enforce a check for existence
of the path in the low level code.

Since client code may indeed not care about existence of the directory,
I leave bpy_resource_path alone and patch the python code to check for
existence. An extra argument that enforces the check might be a better
solution, but I will be leaving this to the UI team.

Minor fix, report missing instead of corrupt icon if file does not exist

Fix OpenGL extension report in system info operator.

Fix compilation on gcc

Fix compilation

Merge branch 'master' into temp_display_optimization

Painstakingly merge branch 'master' into temp_viewport_fx_merge

This branch will not make it into blender, however there are a few things that might be useful here for people who work on the viewport project and could be merged to master:

* Matrix stacks
There is working code for matrix stacks in gpu_matrix.c. Basically it works by overriding all draw calls with a special version that loads our custom stack in OpenGL. To refactor this out, coders should do this in three stages:

1) Inspect all calls to GPUBegin/GPUDraw*/GPURect* and only add one call to upload the matrices once to the relevant uniform of the shader, then substitute with the normal GL call.

2) Remove all hardcoded GL matrices from shaders. Shaders should use generic uniforms/uniform blocks for matrices

3) Do not use GL-style stacks. Actually stacks should be generic and local, not global. We do not care about a global projection/modelview etc matrix, rather we want to add matrix stacks locally in places where they are needed and upload those matrices to our shaders. The code can be generalized to do that by substituting the enum in the gpuMatrix* functions with a GPUMatrixStack* though it will be slightly tedious to do - like everything viewport related.

* GLU removal:

This branch also removes glu completely from blender by providing a helper library that handles drawing of some helper shapes. The way this is handled is not ideal, since it uses the GSOC code that is not the way we want to do things here, however it could be improved and made with a proper design.

There are some graphical glitches still here, but should be fixable.