[FEA] Split batches from parquet that are too large, and try to guess better before decompressing

[revans2]

Is your feature request related to a problem? Please describe.
You might consider this a bug or an enterprise feature. I am fine either way. Occasionally we get really crazy compression ratios on ORC and Parquet files. Right now we have a config to limit the size of the data based off of compressed data input. But if we hit a situation where the compression ratio is really good we can violate the batch size limit, by a lot. Especially if we are reading all of the columns in the file. It would be great if we could.

1. guess at the output size of the data based off of row group metrics (number of rows) and schema. Most of the time we see this is for fixed width types and if we can tell early on that we are going to try to decompress too much data we should limit the amount sent to the GPU early.
2. If we do guess wrong (by a lot, and we need to decide what a lot means) we should split the batch and allow some of the batches to be spillable while we process the first batch.