[BUG] to_orc writes incorrect sum statistics when there's an overflow

[ayushdg]

Describe the bug
cudf.read_orc with certain predicate filters fails in cases where the sum of column values in the column being filtered exceeds int64 limits.

Steps/Code to reproduce bug

import cudf
import numpy as np

np.iinfo(np.int64).max
df = cudf.DataFrame()
df['key'] = [np.iinfo(np.int64).max - 1, 2]
df['val'] = [1,2]
df.to_orc("predicate_overflow.orc")
df = cudf.read_orc("predicate_overflow.orc",filters=[("key", ">", 0)]) 
print(len(df)) # 0

Expected behavior
The returned dataframe doesn't lose valid rows (2 rows in this case).

Environment overview (please complete the following information)

• Environment location: bare-metal
• Method of cuDF install: conda

Environment details
21.10 nightly from today (Aug 27)

Additional context
Some minor debugging indicated that the logic here fails when the col_sum returned from gathering metadata is a negative value because of overflow.

cc:@randerzander

[ayushdg]

Quick update here: Looks like the issue might be with cudf.to_orc and how orc_statistics are written.
I verified for the example above that the raw orc statistics value for sum is a negative number (indicating an overflow).

From the orc specification: if the sum overflows long at any point during the calculation, no sum is recorded., so this seems to be a case where we are incorrectly including the sum statistic within the orc metadata.