[ARM] Add VADDV and VMLAV patterns for v16i16

This adds patterns for v16i16's vecreduce, using all the existing code
to go via an i32 VADDV/VMLAV and truncating the result.

Differential Revision: https://reviews.llvm.org/D85452