Terminal evaluator: draw evaluation RnD

[lrozenblyum]

Terminal evaluator produces 1 for winning, 0 for losing which are the furthest possible points of evaluation.

However providing 0.5 for draw was not ok (as it causes #299).

Our master evaluator doesn't tell at the moment that if evaluation is > 0.5 we shouldn't accept draw and if it's < 0.5 we should accept draw [ IDEAL_CONSTRAINT ] .

Naive implementation of #299 is the first step to do more extensive RnD.

The treatment of accept draw should be revised.
Ideas
[ ] Normalized brain should produce the figures that truly match the IDEAL_CONSTRAINT
[ ] Draw evaluation should be delegated to some another component (contradicts with single point of responsibility for terminal positions)
[ ] ? any other ideas?