Precompute X_t in partial dependence slow mode

In partial dependence fast mode, we precompute `pipeline.transform_all_but_final(X_eval)`, which allows us to determine if a feature will have an impact on model results and, if it won't, use averaged predictions from that X_t as partial dependence values.

We do not do this in fast mode, as we don't have any other need for the precomputed X_t, but if we did have it, we could similarly avoid most of the computation in those partial dependence cases. This issue is to consider whether it is worth it to precompute X_t in slow mode as well. This would add more overhead for all slow mode runs, but would greatly increase performance in some cases, so we should run performance tests and investigate how much more overhead it adds and how often we'd see the benefits.