nv50/ir: normalize cube coordinates after derivatives have been computed

nv50/ir: normalize cube coordinates after derivatives have been computed

In "manual" derivative mode (always used on nv50 and sometimes on nvc0
but always for cube), the idea is that using the quadop instruction, we
set up the "other" quads to have values such that the derivatives work
out, and then run the texture instruction as if nothing were strange. It
pulls values from the other lanes, and does its magic.

However cube coordinates have to be normalized - one of the 3 coords has
to be 1, to determine which is the major axis, to say which face is
being sampled. We were normalizing the coordinates first, and then
adding the derivatives. This is wrong for two reasons:

- the coordinates got normalized by a scaling factor but the
  derivatives didn't
- the result of the addition didn't end up normalized

To resolve this, we flip the logic around to normalize *after* the
per-lane coordinates are set up.

This fixes a bunch of textureGrad cube dEQP tests.

NOTE: nv50 cube arrays with explicit derivatives are still broken, to be
resolved at a later date.

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Cc: "11.1 11.2" <mesa-stable@lists.freedesktop.org>