What this fixes

A missing value (None) in a Json() column is currently stored as the JSON string "null" instead of a real Arrow null. This has two visible consequences:

1. The column reports null_count == 0 even when values are missing, and is_null() returns all False. Anything that relies on Arrow null semantics (filtering on nulls, counting nulls, null-aware Parquet writes) sees no nulls at all.
2. A genuine missing value becomes indistinguishable from a column that holds the literal JSON value null. Both end up as the same "null" string in storage.

Reproducer on current main:

from datasets import Dataset, Features, Json

ds = Dataset.from_dict({"col": [{"a": 1}, None, {"b": 2}]}, features=Features({"col": Json()}))
arr = ds.data["col"].combine_chunks()
print(arr.null_count)        # 0, expected 1
print(arr.to_pylist())       # ['{"a":1}', 'null', '{"b":2}'], expected the middle to be a real null

Root cause

The from_dict path and the on_mixed_types="use_json" path encode each value through json_encode_field in src/datasets/utils/json.py. Its leaf branch ran ujson_loads(example) on the value. For None that raises, so it fell into the except and returned ujson_dumps(None), which is the string "null". The same gap existed in Json.encode_example. On the read side Json.decode_example(None) crashed because it passed None straight to ujson_loads, and Json.cast_storage re-serialized None entries to "null" when it had to rebuild the array.

The fix

None is now preserved as a real Arrow null across all of these entry points:

• json_encode_field and json_decode_field return None unchanged at the leaf.
• Json.encode_example(None) returns None.
• Json.decode_example(None) returns None instead of crashing.
• Json.cast_storage keeps None entries as nulls when it rebuilds the storage array, and the validation sampling skips None so a sampled null does not force a needless re-encode.

Decoded access through ds[:], to_list and to_dict is unchanged, so this does not change observed values for existing users. It only fixes the underlying storage so nulls behave like nulls.

Verification

Added two regression tests in tests/test_arrow_dataset.py (test_json_feature_keeps_none_as_null and test_json_feature_all_none). Both fail before the change and pass after.

Ran on CPU:

python -m pytest tests/test_arrow_dataset.py -k "json or Json or mixed" -q   # 25 passed
python -m pytest tests/features/test_features.py -k "json or round_trip or yaml" -q   # 58 passed
python -m pytest tests/io/test_json.py tests/packaged_modules/test_json.py -q   # 79 passed
python -m pytest tests/io/test_parquet.py -q   # passed
python -m pytest tests/test_arrow_writer.py -q   # 86 passed

Also confirmed by hand that the mixed-None column now reports null_count == 1, that an all-None column is all nulls, that the on_mixed_types="use_json" path keeps nulls, and that a Parquet roundtrip preserves the null. ruff check and ruff format --check are clean on the changed files.