Invert rows and depth in non-vectorized portion of packing (PowerPC).

Chip Kerchnerrequested to merge
ChipKerchner/eigen:invertDepthRows_MR into master

Invert rows and depth in non-vectorized portion of packing for RHS (PowerPC).

This shows up as bad results in the following:

export EIGEN_SEED=1629216664
test/product_syrk_3
test/product_mmtr_3

The previous packing did NOT allow us to know the correct end of a row in some cases and it would pickup incorrect values from the wrong locations.

In the process of fixing this, I simplified the code and added performance improvements (extra rows are now 5X faster and overall 10% gains).

Merge request reports