Investigate and implement further optimizations for Partial Dependence Fast Mode

The following are some ideas for how to potentially further improve performance in partial dependence's fast mode:
1. Allow passing the transformed X in from outside the partial dependence function - this is useful if you’re running partial dependence multiple times for multiple features in X. This way you avoid the `pipeline.transform_all_but_final(X)` call for however many features you're calculating partial dependence for. Note that we need to make sure to not mutate the original X_t or we’d lose the original feature values when updating X_t for every grid value. It's possible that having to copy this dataframe every time negates any performance improvements from precomputing X_t. This should be investigated
2. Avoid passing in the estimator in when fitting the cloned pipeline. The estimator isn't needed when refitting the cloned pipelines, as we're only ever using it to `transform_all_but_final`. Given the performance differences I’ve seen between using the random forest vs linear regression with the same transformers, I think this might provide some performance improvements, though they might not be super noticeable. Also worth noting that after removing the estimator, we would have to just call transform, since transform_all_but_final would then not get the last transformer component.