[VectorCombine] scalarizeExtExtract uses LE lane mapping on BE targets

[uyoyo0]

The scalarizeExtExtract combine packs into i(N*8) and then extracts lanes via lshr/and. The shift amount assumes little-endian ordering which is incorrect for big-endian targets.

This causes miscompiles on big-endian targets such as AIX/PowerPC under -O3 -flto.

[RKSimon]

Reopening - will backport once its had time to stew for a few days in trunk