fix(core): fix decomposition over 1 level to be balanced

[IceTDrinker]

It was brought to our attention that a decomposition over 1 level was not properly balanced and could have a mean value of 0.5 which can be an issue when being fed in an FFT (generating harmonics in the process) this fixes that.

A tests has been added to check on a selected 1 level decomposition that the average of the values returning during decomposition is ~0

cc @agnesLeroy and maybe @guillermo-oyarzun so that you can check how to update this on your end, I can explain the formula in more detail, but normally the comment in the code explains how it works

[IceTDrinker]

I launched the tests in advance and as you can see all green, so the fix is behaving properly

[IceTDrinker]

while checking with @mayeul-zama it was not clear the bit trick was doing exactly what we wanted, will take a look to fix this and potentially do the same on the decompose_one_level function

[IceTDrinker]

tested my bit trick for input values, the bit trick agrees with the plain if, will check the one in the decomposition as well,

[IceTDrinker]

Ok will push a test validating the bit trick, then once this hits the PR GPU team can look at updating their code

[IceTDrinker]

updated the code, the bit tricks are now in a function marked inline(always) and have a dedicated test to verify their output is correct good to review, I'll launch the tests directly with approved

[IceTDrinker]

something going wrong in the pbs 128, need to check what it is

[IceTDrinker]

latest pbs 128 failure (mismatch between two implems we have) is now fixed

[mayeul-zama]

Thanks!