Fix integer overflow in partition scatter_map construction

[wence-]

Description

Although the input partition map may be any valid integral type, the
intermediate scatter map should have the same type as a valid row
index (and the offsets histogram), namely size_type. Previously, the
scatter map was created with the same integer type as the partition
map, which can result in integer overflow, and incorrect results, when
the partition map is a narrow integral type and the input table has
more rows than the width of the type.

On the Python side, this adds validation of the user
partition_map input to ensure that all entries are in range,
avoiding potential out of bounds memory accesses.

Closes #7513.

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[davidwendt]

Approving for C++

[wence-]

Merged trunk so hopefully the unrelated python test fails will be gone.

[ttnghia]

I think the failed python tests are relevant:

/opt/conda/envs/test/lib/python3.9/site-packages/cudf/core/dataframe.py:2290: in scatter_by_map
    partitioned_columns, output_offsets = libcudf.partitioning.partition(
/opt/conda/envs/test/lib/python3.9/contextlib.py:79: in inner
    return func(*args, **kwds)
partitioning.pyx:59: in cudf._lib.partitioning.partition
    ???
/opt/conda/envs/test/lib/python3.9/site-packages/cudf/core/scalar.py:279: in __bool__
    return bool(self.value)
....

[wence-]

I think the failed python tests are relevant:

/opt/conda/envs/test/lib/python3.9/site-packages/cudf/core/dataframe.py:2290: in scatter_by_map
    partitioned_columns, output_offsets = libcudf.partitioning.partition(
/opt/conda/envs/test/lib/python3.9/contextlib.py:79: in inner
    return func(*args, **kwds)
partitioning.pyx:59: in cudf._lib.partitioning.partition
    ???
/opt/conda/envs/test/lib/python3.9/site-packages/cudf/core/scalar.py:279: in __bool__
    return bool(self.value)
....

Ah, I didn't think about nulls in my error checking. Will look tomorrow.

[wence-]

I think the failed python tests are relevant:

/opt/conda/envs/test/lib/python3.9/site-packages/cudf/core/dataframe.py:2290: in scatter_by_map
    partitioned_columns, output_offsets = libcudf.partitioning.partition(
/opt/conda/envs/test/lib/python3.9/contextlib.py:79: in inner
    return func(*args, **kwds)
partitioning.pyx:59: in cudf._lib.partitioning.partition
    ???
/opt/conda/envs/test/lib/python3.9/site-packages/cudf/core/scalar.py:279: in __bool__
    return bool(self.value)
....

Ah, I didn't think about nulls in my error checking. Will look tomorrow.

OK, should be fixed...

[wence-]

/merge