Avoiding an overflow in the LTP quantization

Reduces the XX values to avoid an overflow in silk_VQ_WMat_EC(), even in
floating-point.