Split the `invalidate` method in two phases: in `invalidateCheck` it
just checks if the model diff warrants the execution of the actual
`invalidate` method, which requires fetching the custom invalidation
data from storage. The actual `invalidation` method remains the same.
This should make invalidation faster for storage providers that store
the invalidation remotely.