Normally we don't validate the parameters in transform yet.

This transformer is a bit special because it is stateless since it does not do anything at fit time.
So I see a couple of trouble here with the current implementation:

First, we should inherit from TransformerMixin:

class PatchExtractor(TransformerMixin, BaseEstimator)

Then, we should state programatically that the transformer is stateless by implementing the following method:

    def __sklearn_is_fitted__(self):
        """Return True since PatchExtractor is stateless."""
        return True

and add the associated tag:

    def _more_tags(self):
        return {"X_types": ["3darray"], "stateless": True}

I assume that we could implement these fixes in another PR. Then, it makes sense to validate the parameter in transform since it could be called without calling fit.
In short, the current PR seems fine to me as is.

I opened #24230 to solve the issue.