[BUG] Parquet writer does not apply offset to nullmask

[devavret]

cuDF's parquet writer writes incorrect file when columns have a null mask and an offset.

In [1]: import cudf

In [2]: df = cudf.DataFrame({'a':[1,None,3,4,5]})

In [3]: df2 = df[2:]

In [4]: df2
Out[4]: 
   a
2  3
3  4
4  5

In [5]: df2.to_parquet("sliced.parquet")

In [6]: cudf.read_parquet("sliced.parquet")
Out[6]: 
      a
0     3
1  <NA>
2     5

The above example demonstrates that the validity of the output was written without taking the input's offset into account and hence the element at index 1 in the output is null.

This is possibly an issue with other cuDF writers as well. I haven't tried to repro with ORC, although the code suggests this might be the case.