RVP: Replace staturating multiply-accumulate with non-saturating one to match CMSIS-NN

[PhilippvK]

We use __rv_kmada all over the place for MAC operations. KMADA is a saturating instruction, aka it prevents accumulator overflows.

It turns out that ARM's CMSIS-NN actually uses SMLAD which is a non-saturating (aka. warping) Multiply-Add. To bit bit-accurate in edge cases where an overflow might occur, we should replace the __rv_kmada with __rv_kmda? (Need to double check this!).

However prio is pretty low since we are implementing the outdated RVP draft anyways. The newer draft has dropped saturating accumulating instructions completely (see [email protected]), hence we would have to replace them anyways when migrating to the newer spec (e.g. make use of PM2ADDA.H).