[R] Fix global feature importance and predict with 1 sample.

[trivialfis]

• Add implementation for tree index. The parameter is not documented in C API since we
  should work on porting the model slicing to R instead of supporting more use of tree
  index.

• Fix the difference between "gain" and "total_gain".

Related: #7260 (comment).

[trivialfis]

@hcho3 @hetong007 Please take a look when you are available.

[hetong007]

Just to confirm, with this patch, xgboost won't break EIX/radiant.model, right?

[trivialfis]

I ran devtools::check() on radiant.model and it passed. For EIX, the vignettes failed to build and it doesn't have any test in its repository https://github.com/ModelOriented/EIX .

  installing the package to build vignettes
E  creating vignettes (26.5s)
   --- re-building ‘EIX.Rmd’ using rmarkdown
      [[ suppressing 19 column names 'satisfaction_level', 'last_evaluation', 'number_project' ... ]]
   Warning: ggrepel: 2 unlabeled data points (too many overlaps). Consider increasing max.overlaps
   Quitting from lines 157-165 (EIX.Rmd) 
   Error: processing vignette 'EIX.Rmd' failed with diagnostics:
   non-numeric matrix extent
   --- failed re-building ‘EIX.Rmd’
   
   --- re-building ‘titanic_data.Rmd’ using rmarkdown
   Warning: ggrepel: 4 unlabeled data points (too many overlaps). Consider increasing max.overlaps
   Quitting from lines 81-86 (titanic_data.Rmd) 
   Error: processing vignette 'titanic_data.Rmd' failed with diagnostics:
   non-numeric matrix extent
   --- failed re-building ‘titanic_data.Rmd’
   
   SUMMARY: processing the following files failed:
     ‘EIX.Rmd’ ‘titanic_data.Rmd’
   
   Error: Vignette re-building failed.
   Execution halted
Error in (function (command = NULL, args = character(), error_on_status = TRUE,  : 
  System command 'R' failed, exit status: 1, stdout + stderr (last 10 lines):

[trivialfis]

Same error as you have shared. Rerunning tests with 1.4

[trivialfis]

@hcho3 sorry, pushed a new commit for the fix in prediction leaf, where input is only one sample but we need to return a matrix instead of vector. I rewrote the prediction conditions to mimic the old code exactly and ran tests with those reverse dependencies.

[trivialfis]

@hetong007 I have tested both packages using devtools.

[trivialfis]

I will back port