Logpdf and score characteristicks

[MyznikovFD]

No description provided.

[Desiment]

I guess there can be inconsistency in calculating score function. As far as i remember, the current pipeline for computation works like this:

1. cast to base parametrization;
2. compute
3. return

But, gradients of functions with respect to parameters will be different in different parametrizations (more specifically they will differ by multiplication of gradient on Jacobi matrix of mapping between parametrizations). I do not know how handle this problem correctly, but

1. this at least should be covered in tests and docstrings
2. this is not quite related to this PR, and seems as more sophisticated task to do

Probably need help of @LeonidElkin

[LeonidElkin]

lgtm. There are no special questions about the code. Mathematical correctness requires approval from @Desiment . Don't make six commits, one will be enough

[LeonidElkin]

No need to create another variable if you use it only once in the expression and it is so small