with current code on plot variant example 10 iteration, 3 split, 2 max_overfitting_rounds
IterativeImputer BayesianRidge 0.675916
DecisionTreeRegressor 0.739595
ExtraTreesRegressor 0.696068
KNeighborsRegressor 0.834460

original code: 10 iteration
IterativeImputer BayesianRidge 0.695367
DecisionTreeRegressor 0.764438
ExtraTreesRegressor 0.698408
KNeighborsRegressor 0.834774

Convergence in iterativeImputer is not very clear and current sklearn implementation compares maximum sum of change of all imputed features for each sample between consecutive iterations so different number of imputations for each sample and scale is not used and compares it with biggest value in dataset * tol. A possible consequence of this method is overfitting.

Linking here the related discussion #22053, and a relevant issue #14338.