[ray] Perform preprocessing before hyperopt when possible

[tgaddair]

Currently, we perform preprocessing for every trial separately regardless of whether any preprocessing params are being tuned. This is inefficient, and also seems to cause a deadlock when using RayDatasets with dynamic resource allocation during tuning.

Instead, we can perform preprocessing up front, and then reuse the RayDataset among all the trials. It may be the case that we need to persist to parquet as well, otherwise it may cause issues when multiple trials attempt to run the pipeline simultaneously.