Fix: improve speed of trees with MAE criterion from O(n^2) to O(n log n)

doc/modules/tree.rst

@@ -596,7 +596,7 @@ Mean Absolute Error:
 
     H(Q_m) = \frac{1}{n_m} \sum_{y \in Q_m} |y - median(y)_m|
 
-Note that it fits much slower than the MSE criterion.
+Note that it is 3–6× slower to fit than the MSE criterion as of version 1.8.
 
 .. _tree_missing_value_support:
 

doc/whats_new/upcoming_changes/sklearn.tree/32100.efficiency.rst

@@ -0,0 +1,4 @@
+- :class:`tree.DecisionTreeRegressor` with `criterion="absolute_error"`
+  now runs much faster: O(n log n) complexity against previous O(n^2)
+  allowing to scale to millions of data points, even hundred of millions.
+  By :user:`Arthur Lacote <cakedev0>`

doc/whats_new/upcoming_changes/sklearn.tree/32100.fix.rst

@@ -0,0 +1,6 @@
+- :class:`tree.DecisionTreeRegressor` with `criterion="absolute_error"`
+  would sometimes make sub-optimal splits
+  (i.e. splits that don't minimize the absolute error).
+  Now it's fixed. Hence retraining trees might gives slightly different
+  results.
+  By :user:`Arthur Lacote <cakedev0>`