[AMDGPU] Tune threshold for cmp/select vector lowering

It was set in total vector size while the idea was to limit
a number of instructions. Now it started to work with doubles
and thresholds needs to be updated.

Differential Revision: https://reviews.llvm.org/D80322