Would it be reasonable to avoid np.asarray calls in check_array if the input is array-like?

My guess is that the general answer is "No. The np.asarray call is important to guarantee consistency within scikit-learn's estimators. We strongly value consistency."

The context here is that after the recently merged #11308 , some scikit-learn transformers like RobustScalar that used to work on dask arrays no longer work because they auto-coerce their inputs into numpy arrays. This likely comes up in other situations as well.