Read nullable from Parquet metadata

[madsbk]

In the current implementation, we read the first row of a Parquet file in order to determine which columns within a parquet file are nullable.
This has some significant performance drawback, since we are going to be paying the cost of at least an entire row group in terms of decompression time and memory.

Instead, we should parse the parquet metadata before launching any tasks and extract the nullable information.