[CostModel][X86] Reduce cost of vector selects on SSE2/AVX1 targets

Based off the script from D103695, we were exaggerating the cost of the OR(AND(X,M),AND(Y,~M)) expansion using instruction count instead of effective throughput