Describe the workflow you want to enable

GaussianRandomProjection and SparseRandomProjection should have an inverse_transform() method.

Describe your proposed solution

Simply add the inverse_transform() method, with the appropriate input validation and tests, and return X @ np.linalg.pinv(self.components_.T).

Describe alternatives you've considered, if relevant

N/A

Additional context

My understanding is that every transformer that performs a reversible transformation should have an inverse_transform() method, unless there's a really good reason not to (e.g., if it's terribly inefficient, or there's a better alternative). Was the inverse_transform() left out because it's too inefficient?