Error casting strings with large dates to Timestamp

[ryzhyk]

Describe the bug

Attempting to convert a string with a valid timestamp with a large date in the ISO format (e.g., +10999-12-31T00:00:00) results in:

Error parsing timestamp from '+10999-12-31T00:00:00': error parsing date

To Reproduce

#[test]
fn test_cast_string_with_large_timestamp_to_timestamp() {
    let array = Arc::new(StringArray::from(vec![
        Some("+10999-12-31T00:00:00"),
    ])) as ArrayRef;
    let to_type = DataType::Timestamp(TimeUnit::Second, None);
    let options = CastOptions {
        safe: false,
        format_options: FormatOptions::default(),
    };
    let b = cast_with_options(&array, &to_type, &options).unwrap();
}

Expected behavior
The cast succeeds.

Additional context

I think this issue is similar to #7073, but this one is for timestamps, whereas the other one affected dates.

I ran into this issue when parsing a metadata file of a delta lake table containing large dates. The table was created using Spark SQL i.e., this issue occurs in the wild.

[mbutrovich]

I've actually been debugging a similar issue for DataFusion Comet, and will open a related issue shortly. The issue may stem from the fact that Spark still defaults to writing INT96 for timestamps. In my issue, we read back a Parquet file written with large timestamp values from a Parquet file, and arrow-rs coerces them into a Timestamp(TimeUnit::Nanoseconds, None) by default which cannot represent as large of a date range as an INT96.