Implement `check_lengths`

https://github.com/dask/dask-ml/pull/863/files#diff-970a9a902963a060977ac1912b5a86a5177cebfd468517967f63335bb6d024eaR435-R437 added an empty `check_consistent_lengths`. We should implement a real one.

Unlike scikit-learn, we don't actually want to check the lengths when constructing the graphs. But we do want to check things like that the number of chunks match, that the chunk shapes match if present, divisions, etc.