PandasCursor infers float64 for integer and JSON columns with NULL on managed query result storage

[laughingman7743]

Problem

With managed query result storage, PandasCursor builds its DataFrame from the GetQueryResults rows (AthenaPandasResultSet._as_pandas_from_api(), pyathena/pandas/result_set.py:844-849) and lets pandas infer the dtypes.
A BIGINT column with a NULL becomes float64, so integers above 2**53 lose precision and NULL becomes NaN.
With an S3 result file, the same column is Int64 and keeps the value.

A json column whose values are numbers and NULL becomes float64 on both paths, so large JSON integers also lose precision there.

Expected: the managed path returns the same dtypes and values as the S3 path (Int64 with <NA>), and JSON numbers keep their decoded Python values.

Reproduction

Measured on Athena on master 16aef64:

from pyathena import connect
from pyathena.pandas.cursor import PandasCursor

sql = """
SELECT id, b, j FROM (VALUES
  (1, BIGINT '9007199254740993', json_parse('9007199254740993')),
  (2, NULL, NULL)
) AS t(id, b, j) ORDER BY id
"""
cursor = connect(work_group="<managed work group>", s3_staging_dir="", cursor_class=PandasCursor).cursor()
cursor.execute(sql)
cursor.fetchall()
# [(1, 9007199254740992.0, 9007199254740992.0), (2, nan, nan)]
cursor.as_pandas().dtypes  # id int64, b float64, j float64

With an S3 result file: [(1, 9007199254740993, 9007199254740992.0), (2, <NA>, nan)], dtypes Int64, Int64, float64.

Environment

• PyAthena master 16aef64, Python 3.13.1, pandas 3.0.6, PandasCursor (the async and aio pandas cursors share the result set), a work group with managed query result storage.

Found during the review of #1010 (#934).