Interface to `inject` the gradient conveniently

[nikohansen]

The current inject interface interprets, by default, injected solutions "as directions", however, only by shortening the directional vector solution - es.mean when it is exceedingly long. For the injection of a gradient direction, this is somewhat impractical. The correct injection code, to be called before ask or as a callback to fmin2, would be

es.inject(es.mean - self._random_rescaling_factor_to_mahalanobis_size(gradient) * gradient), force=True)

This looks exceedingly obscure and cumbersome. Hence, we should provide

• the above code as es.inject_gradient(gradient)
• an option which would transform the gradient to an estimated Newton direction using covariance matrix information, for example like es.inject_gradient(gradient, newton=True)
• decide whether and how to handle genotype-phenotype transformations. When the gradient is analytically available, like for boundary handling, this would make sense?

In case of a genotype-phenotype transformation (e.g. with boundary handling and/or fixed variables and/or...), the gradient calculation has to take into account this transformation too, namely es.to_phenotype(es.mean) which is in effect es.gp.pheno(solution, copy=True, into_bounds=es.boundary_handler.repair), because es.mean resides in the genotype space whereas the fitness is computed from the phenotype.

[nikohansen]

We could also name this method .inject_direction rather than .inject_gradient and accompany it with .inject_solution. The second argument could have the same format in both cases, similar to the CSA_clip_length_value option. However, we want additionally distinguish the case when we resample the length.

Why do we even need the CMA_injections_threshold_keep_len option, given we need to call .inject anyways?