Permutation shap sampling estimation + paired sampling

[martinju]

Just a brute force implemention with a basic testing script, to see how it works out.

Based on this small tesst, paired sampling is much better. Permutation seems to be essentially the same as the kernel approach.

TODO:

• Consider whether there are other good reasons to implement the permutation approach
• Do a more comprehensive simulation study and check the performance on indiviual feautres is more balanced than the kernel approach
• If the results suggest the permutation approach is preferable, it must be implemented properly, this is just a brute force thing which is not efficient at all.