Bug in explain_forecast()

[LHBO]

In this PR, we fix 2 main bugs in explain_forecast():

A: Using verbose = "shapley", caused the function to crash.
B: Inconsistencies related to the number of coaltions used and shapley values computed with explain_forecast().

Bug A stemmed from :

1. There was an inconsistency in that internal$iter_list[[iter]]$dt_shapley_est contained the index columns "explain_idx", "horizon" while internal$iter_list[[iter]]$dt_shapley_sd did not. We added the index columns to the latter.

2. When creating a data.table to print with the iterative Shapley values estimates with standard deviations, the code did not take into consideration that the number of rows is not n_explain when doing forecast, but rather n_times x horizon. This is now fixed by using the number of rows in internal$iter_list[[iter]]$dt_shapley_est instead.

3. To be consistent with regular Shapley value, we remove "explain_idx", "horizon" when printing the iterative Shapley value estimates and standard deviation. In regular Shapley we omit "explain_idx".
   Note. It might be a good thing to include "explain_idx" (and "horizon"), but then the print function needs to be rewritten to reflect that "explain_idx" (and "horizon") is not an estimate with a standard deviation.
   Note 2. There is an inconsistency between explain and explain_forecast, where the former use explain_id, while the latter use explain_idx.

Bug B was fixed by defining the number of coaltions to be the number of coalitions used per horizon. We also introduced a new parameter n_sampled_coalitions which for explain_forecast() is the number of coalitions per horizon -2 (zero and full prediction), and for regular explain() it is the number of coalitions. A number of additional minor fixed related to the number of shapley values computed with/without grouping was also added. As a side-effect this also fixed the incorrect verbose printout for explain_forecast().

PS.
print_dt <- as.data.table(matrix(paste(matrix1, " (", matrix2, ") ", sep = ""), nrow = nrow(matrix1)))
produce extra whitespace at the end. Should change to, e.g.,
print_dt <- as.data.table(matrix(paste0(matrix1, " (", matrix2, ")"), nrow = nrow(matrix1))).

[LHBO]

To see that bug no longer occurs, run:

devtools::load_all()
explain_forecast(
 testing = TRUE,
 model = model_arima_temp,
 y = data_arima[1:150, "Temp"],
 xreg = data_arima[, "Wind"],
 train_idx = 3:148,
 explain_idx = 149:150,
 explain_y_lags = 3,
 explain_xreg_lags = 3,
 horizon = 3,
 approach = "empirical",
 phi0 = p0_ar,
 group_lags = FALSE,
 max_n_coalitions = 150,
 iterative = TRUE,
 iterative_args = list(initial_n_coalitions = 10),
 verbose = c("basic", "progress", "convergence", "shapley", "vS_details")
)

[LHBO]

I am confident that the test that fails now are actually correct, in the sense that the output of the old runs were incorrect.
As in the previous code in check_convergence.R, the forecast dt_shapley_sd did not contain the index columns and then max_sd <- dt_shapley_sd[, max(.SD, na.rm = TRUE), .SDcols = -1, by = .I]$V1 removed none column, while dt_shapley_est0[, maxval := max(.SD, na.rm = TRUE), .SDcols = -c(1, 2), by = .I] removed explain_idx and horizon but kept none, as dt_shapley_est had the index columns.

Meaning that @martinju can accept the changes to the test output and merge.