[TEST] improve two-sided test for `base_variable_importance`

[jpaillard]

The class of tests was initially designed for a one-sided test and low p-values. It would be interesting to properly test two-sided alternatives with low one_minus_pvalues_