Adding the separate and surrogate regression method classes

[LHBO]

In this pull request, we add the option to use the separate and surrogate regression method classes discussed in Olsen et al. (2023). We use the packages in the tidymodels framework to specify the models, tune hyperparameters, and fit them.

Key points in this pull request are the following:

1. A detailed vignette that explains the method classes and all points below. It also includes several comparisons with the Monte Carlo-based methods, and the regression-based methods perform better (even without CV). Note that this is not a proper simulation study or anything, but it illustrates that the regression-based methods are very competitive both timewise and accuracy-wise regarding the MSEv criterion.
2. We introduce the separate regression method class to shapr.
3. We introduce the surrogate regression method class to shapr.
4. The regression models (both separate and surrogate) can be any regression model from the parsnip/tidymodels package.
5. Cross-validation allows easy tuning of any parameter in the regression models. The possible hyperparameter values are specified by the user in a data frame or as the output of a user-specified function (in situations where the allowed values change best on the number of features included in the coalitions). The user gets feedback from the CV procedure when verbose = 2.
6. The data can be pre-processed before being sent to the model using recipes. The user can provide a function that modifies the recipe object.
7. Allow the user to create/specify their own regression models that the regression method class is to use.
8. Extend shapr to support explaining any tidymodels fitted using the workflow procedure; see the vignette.
9. The separate regression models can be trained in parallel using future as usual in shapr. The cross-validation procedure of the surrogate regression methods can be parallelized using the same package. The future package also parallelizes the prediction step for both the separate and surrogate regression methods.
10. Made changes to Python code: see list below.

We have also made several changes to the Python version of explain() to make the methods work in Python.

1. Fixed such that we can use regression_separate and regression_surrogate in Python.
2. Fixed indentation.
3. Specify where .utils is loaded from.
4. We added **kwargs to explain() so that we can actually send approach-specific arguments. These must have the form approach_parameter_name in Python, but we added code to translate them to the R syntax approach.parameter_name.
5. Added documentation and some comments.
6. I refactored compute_vS() so that it splits between whether we use an MC-based or regression-based method.
7. As Python and pyr2 do not support tidymodels, we remove all objects related to regression in the internal list before converting them to Python objects and returning them.

[martinju]

Looks good. See some comments. I will look at the vignettes (and recompile them) separately.

[martinju]

First I think we should refer to the pkgdown website for documentation.

What we could do to more easily refer to the ...-parameters in python is to create a new function in R called tripledot_docs or something like that which takes ... as input and where we use @inheritDotParams explain in the documentation, then we can link directly to that function from this documentation (still describing that you need to replace dots by underscores.)

[LHBO]

Added this function and updated the python documentation

[martinju]

So you have moved from 2 spaced indentation to 4? I am Ok with that change, but making that change here makes it hard to spot what are actual code changes and not. I suggest to revert that to ease code review. I am sure there is quick-tool for that.

[LHBO]

PEP 8 standard. I go back to non-standard formatting. We need to determine what to do in the future in another PR.

[LHBO]

I have made all the alterations requested in the previous review, and I have built the vignettes.