[optimizing] float32 and small refactoring

[derVedro]

Using float32 for the computation should speed up calculations a little when dealing with larger arrays. Unfortunately, the results are almost imperceptible and can be considered statistical errors. However another change - the factoring out a subtraction out of score method - give a minor boost for larger crops and a bigger amount of crops (for example for other min_scale, max_scale and scale_step values besides the defaults).