Blender Git Loki

Fix for bug #8865: on mac os x, with certain processors (I'm guessing
Intel Xeon only), doing a lot of mutex locking is really slow. Getting
the image buffer for each texture read then made using more threads
actually slow down the render. Now I've split up the function in two
parts, one parts that checks if the image is available, and another
that does a mutex lock and loading if needed.

Changes quite a lot of code, so hopefully doesn't break stuff, but it
seemed to survive test with rendering a number of frames using all
image types and many threads, though this kind of threading problem
only happens once in a while .. so hard to test for.