Computation simplification with precision reduction

[SuperSashka]

In some computer vision applications it is viable to reduce precision from, for example, float64 to float32/float16 or even to int. Maybe, it would be more faster to compute first guess in, for example, float16 and move to float64 to find finer solution.

Cf. Lazy computation mode #25

[SuperSashka]

We got memory issue with tensors of size (52,125,125)~800k points. Gradients are of size 16GB. It seems that we have to somehow mitigate it. =)

[SuperSashka]

We need to up this issue and take this issue as the next target.

We have to make sure that we can set floating point precision as an argument.