Extend range of arm_mult_q31

[saturn691]

This is the approach used by arm_mult_q7 and arm_mult_q15. The approach currently used would set the LSB to 0 due to the final left shift. My understanding of why that approach was used is because __SSAT(out, 32) doesn't work, because it operates on 32-bit integers, which means it cannot detect that 0x80000000 > 0x7fffffff, as it will interpret this signed.

On devices that support sqdmulh, the compiler can optimise out the core calculation for that function. It must match these semantics, to generate the optimisations.

[christophe0606]

@saturn691 I have not benchmarked but I think this version would be slower. That's why we sometimes have different versions of the same kernel with different accuracy / speed tradedoff.

It would be better as a new function since it is a different accuracy / speed tradeoff and I don't want to change too much the behavior of a function that is already in use.

Your understanding is right : this version is using __SSAT (because it is faster) but it implies that one bit of accuracy has to be lost otherwise __SSAT could not detect the saturation.

[christophe0606]

I have created a new github issue to track this enhancement for future.
I close this PR.
A new function will be required.