feat(model): Add optional memoization to datasets during model training.

[ErikBavenstrand]

By specifying the optional parameter memoized_dataset_cache_size > 0, the corresponding number of datasets will be kept in memory to avoid repeated conversion from vaex to pandas in settings where we perform repeated fitting using the same datasets e.g. hyperparamter tuning. Use with caution and always call clear_load_dataset_cache once completed to clear the cache.