Splits feature_combinations into multiple functions

[nikolase90]

Summarized

• Extract parts of feature_combinations into two new functions called feature_exact and feature_not_exact. All the functions are stored in R/features.R
• Removes observation_weights since it is not used anywhere in the codebase. The function was previously stored in R/shapley.R.
• Updates shapley_weigths so that it handle cases with zero features and m features, i.e. where s = 0 or s = m.
• Moves sampling of features into sampling_features_cpp. This part will only be used if exact = FALSE in feature_combinations().
• Adds tests for functions in R/features.R. See tests/testthat/test-features.R.

Fixes #71 and #72. These issues came up when writing the tests.

Benchmarks

Ran the following chunk using the version from master and nikolai/tests_features

new <- bench::mark(
  new = feature_combinations(m = 15, exact = FALSE, noSamp = 1e4), 
  iterations = 50, 
  filter_gc = FALSE
)
new <- data.table::data.table(new)
knitr::kable(new[, .SD, .SDcols = 1:10])

Current

|expression |   min|  mean| median|   max|  itr/sec| mem_alloc| n_gc| n_itr| total_time|
|:----------|-----:|-----:|------:|-----:|--------:|---------:|----:|-----:|----------:|
|new        | 129ms| 149ms|  144ms| 365ms| 6.704927|        0B|   19|    50|      7.46s|

Master

|expression |   min|  mean| median|   max| itr/sec| mem_alloc| n_gc| n_itr| total_time|
|:----------|-----:|-----:|------:|-----:|-------:|---------:|----:|-----:|----------:|
|old        | 652ms| 770ms|  751ms| 1.03s| 1.29869|        0B|  185|    50|      38.5s|