Excessive memory consumption for large regression dataset

Running NYC Taxi (datasets_v2) with the Default Algorithm (5 batches) results in a large spike in memory consumption. The following memory plot runs just this dataset with 1 trial, measuring the memory every 10 seconds.

As we can see, this plot spikes towards 30GB during the training, which shows the excessive memory consumption we are using and suggests a potential memory leak. I suspect there is some issue with writing the LG results to S3, along with general memory consumption through EvalML.

This issue tracks looking into the memory consumption here and figuring out why we have such large spikes. Would also be useful to look into if this happens for larger classification data as well.