Compute the WB p-values using a strict inequality with a tolerance (a - tol < b) instead of a "a <= b" inequality

This is to be consistent with what the SPM toolbox is doing and is used to circumvent the issue of checking the equality of 2 float values.

I have benchmarked this MR against the SPM version and this removes the small discrepancies that were observed for some WB p-values prior this MR. I have only observed one small discrepancy in a scenario for a voxel and an uncorrected TFCE WB p-value. This can be easily explained by the fact the scale of the TFCE scores is different between FSL and SPM. The latter has already been reported in an issue (#13 (closed)) and I believe we can be happy with this MR. Thus, I am merging straight away.