feat: transformations base primitives

[LeonidElkin]

Base primitives and affine transformation for check. There is also an approximation, but it is very raw for now

[LeonidElkin]

This is a base prototype. What surely needs to get done

• A flyweight pattern for base distributions in transformation tree which will be generated as operations progress
• ArithmeticalMixin for "injecting" operations in any Distribution like class e.g. ParametricFamilyDistribution
• Improve the approximation. Make simple interpolation instead of Chebyshev
• And, of course, test it all

[Desiment]

Summary after on-line call:

We decided to re-consider architecture of the graph (yes, again); from new POV one can see analytical computations as loops with different labels is such graph. This will allow to

• Make choice between analytical computations in complex distributions, such as genhyperbolic
• Allow to control, how characteristics are computed in DerivedDistribution

More specifically, we assume that

1. Analytical calculations inside Distribution now declared as:

dict[GenericCharacteristicName, dict[Label, AnalyticalComputation]]

NOTE: options are still needed, because some of them may take values in continuous domain, or have discrete set of values, but it would be quite sad to redefine them in every transformation; labels and options should be orthogonal in that sense;
2. Graph view considers analytical calculations as loops. The default strategy task is to achieve the first characteristic that have such loops.
3. Since we can have multiple analytical calculations, we need to solve them in a strategy. For now, choice the random one, in the future switch to the preferred-label method.
4. In the future, we'll need a separate strategy for DerivedDistribution (ideally,each transformation will have its own strategy, or one large strategy that observes which operation was performed)

[Desiment]

1. Remove TransformOperationMixin from base ABC distribution, since the latter one can be used not only for numeric distribution, but for e.g. strings; this also should solve cyclic import problem
2. Exploit the fact, that if one works with derived distribution in the form $f(\xi)$ or $g(\xi, \eta)$, when sampling, one can sample $x_1, \ldots, x_n$ from $\xi$ and compute $f(x_1), \ldots, f(x_n)$ (and sample $y_1, \ldots, y_n$, from $\eta$, then $g(x_1, y_1), \ldots, g(x_n, y_n)$ is sample from $g(\xi, \eta)$)
3. Interpolation should be characteristic-based rather that distribution based. I.e. one can have interpolation method for PDF, and for CDF, and for PPF, and then approximation consumes those specific interpolation methods; Probably can move out from transformations