The tolerance is later multiplied by 10**DECIMAL_PLACES, so if we do
it here as well we get a ridiculously high threshold.

Luckily, the validation still passes after fixing this, so we dont need
to do any new measurements.

It seems that either EMA or the RAPL counter had an overflow

Must have missed this one for some reason

Using 1024x512 means that CA, NAS-FT and SRAD all use the same field
size.

With 128x128x32, 20 iterations completes way too quickly, so raise them
significantly.

The binary created by flang sometimes randomly runs into a segmentation
fault. Since the same code works fine with gfortran, I am inclined to
blame the WIP flang compiler for it. If it happens during a measurement
run, just try again until it works.

... to show that python is unsuitable.