lavr: add x86-optimized functions for mixing 2 to 1 s16p with float coeffs