ReadPixels with GL_STENCIL_INDEX does too much packing
This is not too important, as stencil ReadPixels performance isn't critical, but...
ReadPixels with GL_STENCIL_INDEX does MapRenderbuffer on a Z24_S8 buffer, which ends up combining the separate Z24 and S8 buffers into a packed Z24_S8 buffer...and then we re-split it to take the stencil. This is unnecessary work. We may need to update the MapRenderbuffer API to take a "I just want the depth aspect, or stencil aspect" flag.