Fixed-Point Helium Complex Multiplication saturation sensitivy

[FabKlein]

Combining vqdmlsdh and vqdmladhx for running complex multiplications operations on full scale vectors is likely to saturate
It would be safer to downscale the inputs prior to the multiplication.

arm_cmplx_mult_cmplx_q15/q31 would be modified the following way

{
            vecSrcA = vld1q(pSrcA);    pSrcA += ELTS(vecSrcA);
            vecSrcB = vld1q(pSrcB);    pSrcB += ELTS(vecSrcB);

           /* downscale inputs */          
            vecSrcA >>= 1;
            vecSrcB >>= 1;
         
            vecDst = vqdmlsdhq(vuninitializedq(vecSrcA), vecSrcA, vecSrcB);
            vecDst = vqdmladhxq(vecDst, vecSrcA, vecSrcB);

           vst1q(pDst, vecDst);           pDst += ELTS(vecDst);
}

In this case downscaling of the result before storing must be removed

Other locations involving vqdmlsdh / vqdmladhx pairs should be verified as well