Fix PR 17368: disable unnecessary vector multiplication distribution for square of add/sub on ARM

[Tim Northover]

> Please help to review the patch

The change itself looks almost trivially right, but that test seems
excessively complicated to me; it's really not obvious what the key
point is if it ever breaks.

Is there a reason why you can't define @no_distribute to take (<8 x
i8>, <8 x i8>) and return <8 x i8>?

Cheers.

Tim.