Preserve float columns when JSON loader uses field=

[LeSingh1]

Closes #6937.

When load_dataset("json", data_files=..., field="data", ...) is used, columns whose values are all integer-valued floats ([0.0, 1.0, 2.0]) get silently coerced to int64. Repro:

import tempfile, json
from datasets import load_dataset

with tempfile.NamedTemporaryFile(mode="w", suffix=".json", delete=False) as f:
    json.dump({"data": [{"col": v} for v in [0.0, 1.0, 2.0]]}, f)

ds = load_dataset("json", data_files=f.name, field="data", split="train")
print(ds.features)  # before: {'col': Value('int64')}
                    # after:  {'col': Value('float64')}

The underlying cause is pd.read_json(..., dtype_backend="pyarrow") (tracked upstream at pandas-dev/pandas#58866). The path that hits it is the one introduced in #6914 to preserve column insertion order. The dataset-viewer statistics regression in the issue is a direct consequence.

This PR replaces the pandas.read_json -> pa.Table.from_pandas path in the field= branch with a small _arrow_table_from_field helper that builds the Arrow table directly from the already-parsed Python object. PyArrow's own JSON inference preserves float64, and CPython dict iteration order gives us the #6914 column-insertion-order invariant for free (no pandas roundtrip needed).

The helper handles:

• list of dicts (the common case): collect keys in insertion order, build a column-major dict
• list of scalars: wrap in a single-column table named after the configured feature, falling back to "text" (matches the prior df.columns == [0] rename)
• dict of lists (column-major payload): pass through to pa.Table.from_pydict
• empty list with features supplied: emit empty columns matching the configured feature names so downstream _cast_table aligns

The other JSON loading path (raw JSON Lines via paj.read_json) does not have this bug and is unchanged.

There is an older dormant PR for this issue (#7635 from June 2025, no reviews). I went a different direction because that PR scans the resulting DataFrame and re-casts float-looking int columns back to float after the fact. By that point pandas has already converted the values to Python ints, so the isinstance(x, float) check there does not actually detect them, and the scan adds an O(rows) Python pass to a hot path. Sidestepping pandas entirely is simpler and faster.

Tests in tests/packaged_modules/test_json.py:

• test_json_field_path_preserves_float_columns: the exact JSON loader implicitly coerces floats to integers #6937 repro, asserts col stays float64
• test_json_field_path_preserves_float_columns_alongside_ints: list-of-dicts payload with mixed int / float / int-valued-float columns, each type preserved
• test_json_field_path_preserves_float_columns_with_dict_of_lists: dict-of-lists field payload
• test_json_field_path_preserves_column_order_with_list_of_dicts: column insertion order check (sanity vs the Preserve JSON column order and support list of strings field #6914 invariant)

pytest tests/packaged_modules/test_json.py shows 33 passed, 0 failed (29 pre-existing + 4 new).