Reading large parquet datasets

[cdbethune]

For image explanations, we needed to test against big earth single when unpooled and prefeaturized, which ends up ends up being 10K rows x 32K images. The parquet loading library we use seems to be aggressively chewing through RAM, and seems to be leaking memory during the process of loading. A dataset that is about 5GB ends up occupying around 15GB of RAM after being loaded. It is also much slower than the parquet reader that is part of PANDAS, with performance being reduced linearly as it reads columns.