NPOT texture sampling is slightly slower, but with OpenGL we were
already treating every mipmapped texture as NPOT due to requiring
padding at the 1x1 level for renderability. Seamless cube maps are
also NPOT due to the border. Furthermore, the vector shifts by scalar
required for POT texel address calculation would require 128-bit
values when we deprecate MMX.

Change-Id: Ie2e68f632bea7c6f3e599015c14be50392ea7e9a
Reviewed-on: https://swiftshader-review.googlesource.com/8248Reviewed-by: Nicolas Capens <capn@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>
Tested-by: Nicolas Capens <capn@google.com>