Fixed test_similarityAccuracy to have better diffing and some tolerance on the bits checked

The standard output is no more used, replaced by some custom checking and diffing code allowing to select the tolerance on the mantissa of the double numbers at the bit level

However, the tolerance has to be put to 10 bits for the tests to pass in debug (reference being the opt version). This sounds a bit high and one should understand why we are so unstable numerically