The non-vectorized SQDMULL implementation was wrong for 32-bit
inputs.  It incorrectly checked one of the operands to see if the
value would overflow before doubling it, not the result of the
initial multiplication.  It now matches the 16-bit operand version,
and also matches hardware.  A test has been added for the scalar
form of the function, testing a range of values that will saturate
when multiplied and doubled.  This set of test vectors was produced
on an ARMv9 machine (Google Cloud box), failed tests on x86 with the
existing code, and passes with the modified code.