This should more closely reflect what applications actually use.
The basic idea here is that it is better for apps to run out of
descriptors before running out of sets and thus reduce the
overall memory footprint of the application.
Fixes a few bottlenecks that were encountered in the Cascading Shadow
Maps demo from the Microsoft SDK. Performance is now slightly better
than wined3d with CSMT, MESA_NO_ERROR and mesa_glthread enabled.