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.

This reverts commit b6bd5f24.

That way we spend more time using all cores instead of doing single-
threaded serialization of CSV files.

That way we don't multiply any potential errors