Change quantification of matrix multiplication

* increase coeff +2 bits
* add saturation and proper rounding on result (sat 7 bits, round
  15 bits)
* In scale with Libera implementation of MatMult