Allow for larger LOD difference and fix another offset bug
Certain valid implementations were failing couple of
mipmap.cube.projected tests and the reason appears to be that sampling
analytical derivatives around pixel neighborhood may give too tight LOD
bounds, especially if one or more samples fall into different faces.
This change compensates for the difference by reducing assumed LOD
precision down to 3 bits.
In addition this change fixes another LOD offset issue missed by
previous CL.
Bug:
19170879
Change-Id: I5c154294f2b47222f6910e89dfdd372d7d8233bd