ref: https://github.com/scikit-learn/scikit-learn/pull/22986/files#r862344847

Along the way we've had discussions on whether we should raise if a metadata is requested but not provided.

Mirroring the existing behavior, SLEP006 and the implementation won't raise in the following case:

pipeline = make_pipeline(LogisticRegression().set_fit_request(sample_weight=True)).fit(X, y)

Should we allow a strict mode, either globally or on the estimator level, which would make the above code raise?