Do not round real-valued bounds

[MLopez-Ibanez]

The problem is apparently that this round() call sometimes results in irace seeing parameter bounds that are too wide; e.g. when the actual lower bound is 1e-5, then irace with the default readParameters digits value ends up calling the function with value 0. I am doing quite a lot of experiments currently and observed this often; maybe the mutator in irace tries the exact boundary value with nonzero probability? (Another example is lower bound log(2) (0.6931472) -> irace tries 0.6931, which is too low.)

We also accept some tolerance in our real-valued search space components, but that is just much tighter than four digits. I think we should call irace in a way that avoids rounding as much as possible.

Originally posted by @mb706 in mlr-org/bbotk#191 (comment)