Blend packed 16-bit integers from a and b within 128-bit lanes using 8-bit control mask imm8, in each of the two lanes. Note: this is functionally equivalent to two _mm_blend_epi16.