gallivm: do per-pixel cube face selection (finally!!!)

gallivm: do per-pixel cube face selection (finally!!!)

This proved to be tricky, the problem is that after selection/mirroring
we cannot calculate reasonable derivatives (if not all pixels in a quad
end up on the same face the derivatives could get "randomly" exceedingly
large).
However, it is actually quite easy to simply calculate the derivatives
before selection/mirroring and then transform them similar to
the cube coordinates (they only need selection/projection, but not
mirroring as we're not interested in the sign bit, of course). While
there is a tiny bit more work to do (need to calculate derivs for 3
coords instead of 2, and additional selects) it also simplifies things
somewhat for the coord selection itself (as we save some broadcast aos
shuffles, and we don't need to calculate the average vector) - hence if
derivatives aren't needed this should actually be faster.
Also, this has the benefit that this will (trivially) work for explicit
derivatives too, which we completely ignored before that (will be in a
separate commit for better trackability).
Note that while the way for getting rho looks very different, it should
result in "nearly" the same values as before (the "nearly" is only because
before the code would choose the face based on an "average" vector and hence
the derivatives calculated according to this face, where now (for implicit
derivatives) the derivatives are projected on the face selected for the
first (top-left) pixel in a quad, so not necessarly the same face).
The transformation done might not quite be state-of-the-art, calculating
length(dx,dy) as max(dx,dy) certainly isn't neither but this stays the
same as before (that is I think a better transform would _somehow_ take
the "derivative major axis" into account so that derivative changes in
the major axis wouldn't get ignored).
Should solve some accuracy problems with cubemaps (can easily be seen with
the cubemap demo when switching wrapping/filtering), though we still don't
do seamless filtering to fix it completely (so not per-sample but per-pixel
is certainly better than per-quad and already sufficient for accurate
results with nearest tex filter).

As for performance, it seems to be a tiny bit faster too (maybe 3% or so
with cubemap demo). Which I'd have expected with nearest/nearest filtering
where this will be less instructions, but the difference seems to actually
be larger with linear/linear_mipmap_linear where it is slightly more
instructions, probably the code appears less serialized allowing better
scheduling (on a sandy bridge cpu). It actually seems to be now at least
as fast as the old path using a conditional when using 128bit vectors too
(that is probably more a result of testing with a newer cpu though), for now
that old path is still there but unused.
No piglit regressions.

Reviewed-by: Jose Fonseca <jfonseca@vmware.com>