[VE] Inefficient code for icmp i128

[efriedma-quic]

As discussed on https://reviews.llvm.org/D151358, the VE backend generates inefficient code for something like the following:

target triple = "ve"
define i64 @f(i64 %x, i64 %y, i128 %a, i128 %b) {
  %c = icmp ugt i128 %a, %b
  %d = select i1 %c, i64 %y, i64 %x
  ret i64 %d
}

This currently generates 8 instructions, but it can be done in 4 instructions. We currently use 4 extra instructions to translate the result of cmpu into a boolean.

With D151358, this also impacts the efficiency of i128 smin/smax/umin/umax.