Change ridge's cholesky solver from scipy.linalg.solve(assume_a='pos') to scipy.linalg.cholesky

[agoscinski]

In PR #31 we using scipy.linalg.solve with argument assuma_a='pos', to reproduce this with torch however we should use scipy.linalg.cholesky so we can replace it later with equistore operations. This needs a bit additional work (how some additional postprocessing, checking if the scipy and torch solvers bring similar accuracy), so we deal with this in a new PR.