Variable importance for causal forests: breaking down the heterogeneity of treatment effects