[X86] LowerFunnelShift - improve handling of vXi8 constant splat funnel shifts

This patch moves the promotion to vXi16 shifts and the upper/lower bit masking into LowerFunnelShift for targets that have a bit-select instruction (XOP's VPCMOV and AVX512's VPTERNLOG).

This prevents the regressions in #89115 due to the masking of ((X << V) | (Y >> (8-V))) vXi8 shifts.