[X86] Fix the cost of v16i1->v16i16 sext/zext on avx targets.

Previously we were hitting the scalarization case in the default
implementation.