Avoid failing optimization when all pipelines are evaluated with errors

Created by: gkirgizov

The problematic behavior is in method evaluate_with_cache in evaluation.py:

    def evaluate_with_cache(self, population: PopulationT) -> PopulationT:
        reversed_population = list(reversed(population))
        self._remote_compute_cache(reversed_population)
        evaluated_population = self.evaluate_population(reversed_population)
        self._reset_eval_cache()
        if not evaluated_population and reversed_population:
            raise AttributeError('Too many fitness evaluation errors. Composing stopped.')
        return evaluated_population

Failing right away seems unnecessary. I have following thoughts:

Making return type Optional seems more logical, as Optimizer can recover from situation in most cases.
For example, Optimizer could mutate (in a different way) last successfully evaluated population.
Yet the bad case still must be considered: when it's the very first population, or when we have failed too many times in a row.

The task is to design a solution with these points in mind, so that Optimizer could do something in case when evaluation goes wrong.

It's a more general version of the issue https://github.com/nccr-itmo/FEDOT/issues/767 , because this error sometimes appears in several tests, and in principle can arise anytime due to stochastic nature of optimization.