Current custom lowering of truncate vector handles a source of up to 128 bits, but that only uses one of the two shuffle vector operands. Extend it to use both operands to handle 256 bit sources.
Differential Revision: https://reviews.llvm.org/D68035
