Complex-value parsing loses values (typed array(json), native arrays) and PandasCursor NULL TIME differs by result storage

[laughingman7743]

Problem

Three defects in result value conversion, found by the independent reviews of #1033.
Measured on 2026-10-04 on master 9a373fd and on the #1033 branch.

1. A typed array(json) value whose first element is not a string can fail to parse

TypedValueConverter._convert_typed_array() (pyathena/parser.py) decides whether to try json.loads() from the first few characters of the value (a quote in value[1:10], or a leading [{, [null, [[).
Otherwise it uses the native-format splitter, which splits a JSON string element at its commas.

cursor.execute(
    """SELECT ARRAY[json_parse('123456789'), json_parse('"a,b"')] AS v""",
    result_set_type_hints={"v": "array(json)"},
)
cursor.fetchone()  # JSONDecodeError: Unterminated string starting at: line 1 column 1 (char 0)

Without the type hint, the same value converts to [123456789, 'a,b'].
On master, the same array with the string first also failed; #1033 fixes that order (its JSON elements are re-encoded with json.dumps()), but not this one.
The same check applies to arrays nested in maps and rows.

2. PandasCursor returns NaT for a NULL TIME from a result file and None from GetQueryResults

SELECT * FROM (VALUES CAST('12:34:56' AS TIME), NULL) AS t(v):

Result storage	fetchall()	as_pandas()["v"]
S3 result file	[(time(12, 34, 56),), (NaT,)]	[time(12, 34, 56), NaT]
managed (GetQueryResults)	[(time(12, 34, 56),), (None,)]	[time(12, 34, 56), None]

The result-file path parses time columns with parse_dates and _trunc_date()'s .dt.time, which keeps NULL as NaT; the fallback builds the column from DefaultTypeConverter values.
After #1033, time with time zone columns use a converter in both modes and return None.

3. Native-format arrays lose empty and space-only strings and split elements at commas

Measured on master abc99b0 (2026-10-04). The text is the same from GetQueryResults and from the CSV result file, and the typed (array(varchar)) and untyped conversions agree:

Query value	Athena text	Converted
ARRAY['x', '', 'y']	[x, , y]	['x', 'y']
ARRAY['x', '']	[x, ]	['x']
ARRAY['x', ' ', 'y']	[x, , y]	['x', 'y']
ARRAY['a,b', 'c']	[a,b, c]	['a', 'b', 'c']
ARRAY['']	[]	[]
ARRAY['null', 'NULL']	[null, NULL]	[None, None]
MAP(ARRAY['k', ''], ARRAY['', 'v'])	{=v, k=}	{'': 'v', 'k': ''}
CAST(ROW('', 'x') AS ROW(a varchar, b varchar))	{a=, b=x}	{'a': '', 'b': 'x'}

The array splitter (_split_array_items() and the native branch of _convert_typed_array() in pyathena/parser.py) splits at every comma and drops empty and whitespace-only items.
Athena separates elements with ", ", so splitting at ", " and keeping empty items would restore the first four rows.
ARRAY[''] versus ARRAY[], the text 'null' versus NULL, and elements that contain ", " cannot be told apart from the text.
Maps and rows already keep empty strings.

Expected

1. A typed array(json) value converts as JSON regardless of the type of its first element.
2. PandasCursor represents a NULL TIME the same way in both result storage modes.
3. Native-format arrays keep empty and whitespace-only elements and do not split elements at a comma that is not followed by a space, where the text allows it.