Fix gcc-5 packetmath_12 bug.
There seems to be a gcc-5 bug that's causing data1 within
packetmath_minus_zero_add() to be filled with
(-0, 0), (-0, NaN)when optimizations -O2 or higher are on. This is before any packet
operations are called. Printing the values causes the bug to disappear. I've
double-checked we're not running into any aliasing bugs (the cast from
std::complex<double>* to double* is legal, and using c++ casts
doesn't fix the issue). Compiling with -fno-strict-aliasing does not
solve the issue, so seems to be related to something else. The test works
with gcc-6 and later, and all other compilers/versions.
Initializing the memory to zeroes causes the bug to disappear.