Error: ' "what" must be a function or character string ' with XGBoost

Hi
I have fitted an XGBoost model by transforming a data frame (with both features and target feature) to a dgCMatrix using the sparse.model.matrix function from the "Matrix" package. (cvd_incident is the target feature, complete_train_men is the data frame).

`complete_train_men_sparse = sparse.model.matrix(cvd_incident ~., data = complete_train_men)[,-1]
`

I recieve an error message when trying to create an ALE-plot using the FeatureEffect$new function. I tried following the tip for the pred.function in #29 with some modifications, as my model "boostmod_men" is trained on a dgCMatrix instead of an xgb.DMatrix.

```
pred.fun = function(model, newdata){

previous_na_action <- options('na.action') #store the current na.action
options(na.action='na.pass') #change the na.action

newData_x = sparse.model.matrix(cvd_incident ~., data=newdata)[,-1]

options(na.action=previous_na_action$na.action) #reset the na.action

results <- predict(model, newData_x)

return(results)
}
```
To verify that the function works correctly, I try using it to predict the test set and compare with previously computed values in 'pred_test_men':

```
test<- pred.fun(model = boostmod_men, newdata = complete_test_men)

head(test)
head(pred_test_men)

#the two prints the same values for the first 5 observations in test set, so it works fine
```

Now to where the problem arises:

```
mod <- Predictor$new(model= boostmod_men, data = complete_train_men, pred.fun(boostmod_men, complete_train_men))
eff = FeatureEffect$new(mod, feature = "age")
```

The first line runs without issues, but the second one gives the following error message

> Error in do.call(predict.fun, list(model, newdata = newdata)) :
> 'what' must be a function or character string

Do you know what causes this, and how to fix it?

Start by reading the Predictor$new and FeatureEffect$new entry points and compare the arguments passed in the report with the documented prediction-function API. Reproduce the error using the supplied sparse.model.matrix and pred.fun examples, then verify the expected function argument and confirm that FeatureEffect$new completes without the do.call error.

Written by the indexing model from the issue text.