Fix overflow when decoding durations with large day counts - #1320

Decoding an ISO 8601 duration caps the integer part of each segment at 247. That is enough for hours, minutes and seconds but not for days, since 86400 * 247 does not fit in int64. Day counts from 106751991167301 up to 2**47 wrap around, pass the range check, and then either decode to an unrelated value or raise a bare OverflowError:

>>> msgspec.json.decode(b'"P106751991167301D"', type=datetime.timedelta)
datetime.timedelta(days=-579025221, seconds=61184)
>>> msgspec.json.decode(b'"P140737488355328D"', type=datetime.timedelta)
OverflowError: days=-1158050442; must have magnitude <= 999999999

Smaller out-of-range values like "P1000000000D" already raise ValidationError: Duration is out of range, as the parser's doc comment and test_decode_timedelta_out_of_range expect. This PR caps the integer part at MS_TIMEDELTA_MAX_SECONDS instead. Anything larger is out of range for every unit, and 86400 * x always fits in int64 under the new cap, so the existing range check rejects these inputs too. JSON, MessagePack and convert share this parser.

• Ran: pytest tests/unit/test_common.py -k TestTimeDelta. The 4 new cases fail on main for json and msgpack (8 failures, DID NOT RAISE or OverflowError); with the fix, 200 passed.
• Ran: pytest tests/unit: 6591 passed, 157 skipped.
• Ran: ruff check, ruff format --diff, codespell: clean.
• Checked: boundary values such as P999999999DT86399.999999S and PT86399999999999S decode the same as before. About 200k random durations either match timedelta(seconds=...) or raise ValidationError.