Fix Parquet double reads to properly account for null values

[bmcdonald3]

In the Parquet taxi cab dataset, we were hitting end of stream errors when reading in null values and this was identified by Tess in #3087.

To fix this, we switch to reading in single values in the case where nulls are allowed so that we can check the definition level before inserting a value into the Parquet array.

Potential optimizations are an allow_nan flag to avoid the check and then, additionally, we could modify the previous implementation by reading the definition level and values into temporary vectors and then only assign the value if the value is non-null.

The reason it didnt' work before was that it was bsaed off of batchSize to loop over the values, but batch size doesn't actually guarantee how many values were read, due to row group intricacies.

[stress-tess]

Way simpler and fixes the bug I was seeing! Thanks for take care of this so quickly!