Remove maximum assumptions about NaN/inf values for minifloat configurations

[fabianandresgrob]

Our current minifloat support assumes that the NaN and inf values for that specific format are also the maximum representable values of the given minifloat format, i.e. they cannot be in the middle of the possible range but must be on the high end. This currently allows for an easier implementation of clamping values greater than these special values to exactly them, but definitely limits user options for a customized minifloat format. Therefore, we should remove that assumptions and use a more general approach.