[LegalizeTypes] Don't pad cttz_elts_zero_poison with ones when widening op (#206705)
We only need to pad the widened lanes with ones to handle the case for
an all zeroes input. But for cttz_elts_zero_poison, this is already
poison.
The RISC-V scalable vector test can't be precomitted because it crashes
otherwise trying to lower a get_active_lane_mask from
getMaskFromElementCount.
Also while we're here, switch to using TLI.getTypeToTransformTo to be
consistent with other `WidenVecOp_*` implementations.