st/glx: fix depth/stencil format selection code
This fixes a pretty big performance regression caused by commit
3475e88442c16fb2b50b903fe246b3ebe49da226.
When the user does not request a stencil buffer it's important that we
don't use a depth/stencil format (or at least make it our last choice).
If the user calls glClear(GL_DEPTH_BUFFER_BIT) when we have a combined
depth/stencil buffer, that causes us to hit the clear_with_quad() path
which can be much, much slower than calling pipe_context::clear().
Also, try to use a shallower depth format before a deeper one.