MAINT refactor feature_perturbation in Tree explainers

[glemaitre]

The switch between "interventional" to "tree_path_dependent" being silent is quite surprising.

Instead, I would propose to have an "auto" method that would switch to the method depending if the background dataset provided.

Then when providing a specific algorithm, we will in the future raise an error if the conditions are not met to run the algorithm. For the moment, I propose to raise a FutureWarning to ask the user to change to "auto" if this is really what they want (and therefore keep the code backward compatible).

I find this way more explicit, knowing that "tree_path_depedent" might have an important impact on the symmetric axiom expected by the users.

[codecov]

Codecov Report

Merging #2624 (75fa464) into master (45b85c1) will increase coverage by 0.03%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master    #2624      +/-   ##
==========================================
+ Coverage   51.61%   51.64%   +0.03%     
==========================================
  Files          91       91              
  Lines       13146    13153       +7     
==========================================
+ Hits         6785     6793       +8     
+ Misses       6361     6360       -1     

Impacted Files	Coverage Δ
shap/explainers/_tree.py	69.61% <100.00%> (+0.25%)	⬆️
shap/explainers/_additive.py	33.33% <0.00%> (-0.82%)	⬇️
shap/plots/_force.py	63.36% <0.00%> (+0.43%)	⬆️

📣 Codecov can now indicate which changes are the most critical in Pull Requests. Learn more