[Python] Legacy dataset can't roundtrip Int64 with nulls if partitioned

[asfimport]

If there is partitioning and the column has nulls, Int64 columns may not round trip successfully using the legacy datasets implementation. 

Simple reproduction:

 ```python

import pyarrow as pa
import pyarrow.parquet as pq
import pyarrow.dataset as ds
import tempfile

table = pa.table({
'x': pa.array([None, 7753285016841556620]),
'y': pa.array(['a', 'b'])
})

ds_dir = tempfile.mkdtemp()
pq.write_to_dataset(table, ds_dir, partition_cols=['y'])

table_after = ds.dataset(ds_dir).to_table()
print(table['x'])
print(table_after['x'])
assert table['x'] == table_after['x']

```Java

[
  [
    null,
    7753285016841556620
  ]
]
[
  [
    null
  ],
  [
    7753285016841556992
  ]
]

Reporter: Will Jones / @wjones127

Related issues:

• [Python] Long-term fate of pyarrow.parquet.ParquetDataset (is related to)

_{Note: This issue was originally created as ARROW-15725. Please see the migration documentation for further details.}

[asfimport]

Antoine Pitrou / @pitrou:
Very short but likely explanation:

>>> int(float(7753285016841556620))
7753285016841556992

I suspect the legacy partitioning code treats all numeric keys as floating-point for convenience, but this ends up truncating some very large int64 keys. I'm not sure we want to fix it at this point. @jorisvandenbossche