Implement unique sampling in feature_not_exact

[martinju]

When providing n_combinations to shapr, sampling of features combinations are performed unless a n_combinations is larger than 2^m (number of unique samples), for which all combinations are used instead. Currently, n_combinations specifies the number of samples being performed, not how many unique ones should be used in the end (and that is what matters from a memory/runtime perspective). A drawback of this is that if you have say 13 features and set n_combinations = NULL (to use exact) you get all 8192 unique combinations, while if you set n_combinations = 10000 (our recommendation for larger values of m), you get just 2000-3000 unique features.

Suggesting to change the behavior of n_combinations to represent the unique number of features to sample. This can be achieved by first sampling a larger number of combinations and then cut-off when reaching n_combinations unique combinations. To ensure we sample enough combinations, we should use a while loop to repeat this until n_combinations of unique samples are obtained.

While this could increase runtime for the shapr function a little bit, this part of the code is fast anyway.

[JensWahl]

The functions feature_exact/feature_group and feature_not_exact / feature_group_not_exact (see #277) are very similar and reuse a lot of the same code. We could probably combine them into two functions with an additional group argument?