[Patch AArch64] Use 128-bit vectors when autovectorizing 16-bit float types