Implement predictable `steady_state` genetic scheme

Possible problem #1: Currently each iteration `Selection` is applied twice -- at the beginning of training loop & in `Inheritance` operator when used with `GeneticShemeTypesEnum.steady_state`. This seems unneccessary. Maybe if we remove it, algorithm will be faster.

Possible problem #2: when we have duplicate individuals -- they get higher chance to be selected (see the [code](https://github.com/aimclub/GOLEM/blob/09bd64e52dd70118c994b5e2cb9bb347ad8d9a40/golem/core/optimisers/genetic/operators/inheritance.py#L42)). Probabilistically we get more duplicate individuals further and further at each iteration. Related to #89

Possible enhancement #3: implement `steady_state` scheme as described in classical text-books (Dan Simon, $8.5). Now it doesn't follow standard logic.

Problems must be fixed with experiments showing what changed.