Fix empty offset index for all null columns (#4459)

[tustvold]

Which issue does this PR close?

Closes #4459 (fixes a bug introduced by #4389 in 42.0.0)

Rationale for this change

What changes are included in this PR?

Are there any user-facing changes?

[alamb]

cc @AdamGS

[alamb]

I reviewed the code and test in this PR carefully and it looks good to me

I will also attempt to verify it fixes out actual data problem as well using some internal test data

[alamb]

I verified that without the code change in this PR, this test fails like this:

failures:

---- arrow::arrow_writer::tests::test_writer_all_null stdout ----
thread 'arrow::arrow_writer::tests::test_writer_all_null' panicked at 'assertion failed: `(left == right)`
  left: `0`,
 right: `1`', parquet/src/arrow/arrow_writer/mod.rs:2637:9
stack backtrace:
   0: rust_begin_unwind
             at /rustc/90c541806f23a127002de5b4038be731ba1458ca/library/std/src/panicking.rs:578:5
   1: core::panicking::panic_fmt
             at /rustc/90c541806f23a127002de5b4038be731ba1458ca/library/core/src/panicking.rs:67:14
   2: core::panicking::assert_failed_inner
             at /rustc/90c541806f23a127002de5b4038be731ba1458ca/library/core/src/panicking.rs:274:17
   3: core::panicking::assert_failed
             at /rustc/90c541806f23a127002de5b4038be731ba1458ca/library/core/src/panicking.rs:228:5
   4: parquet::arrow::arrow_writer::tests::test_writer_all_null
             at ./src/arrow/arrow_writer/mod.rs:2637:9
   5: parquet::arrow::arrow_writer::tests::test_writer_all_null::{{closure}}
             at ./src/arrow/arrow_writer/mod.rs:2615:31
   6: core::ops::function::FnOnce::call_once
             at /rustc/90c541806f23a127002de5b4038be731ba1458ca/library/core/src/ops/function.rs:250:5
   7: core::ops::function::FnOnce::call_once
             at /rustc/90c541806f23a127002de5b4038be731ba1458ca/library/core/src/ops/function.rs:250:5
note: Some details are omitted, run with `RUST_BACKTRACE=full` for a verbose backtrace.

[alamb]

I have verified that using master and 42.0.0, when I a parquet file of our internal data, the file errors on read with Parquet error: StructArrayReader out of sync in read_records, expected 0 skipped, got 11:

$ datafusion-cli -c "SELECT xxx, time FROM '/tmp/foo.parquet' WHERE 1684850057953220316 <= time::bigint";
DataFusion CLI v27.0.0
Arrow error: External error: Arrow: Parquet argument error: Parquet error: StructArrayReader out of sync in read_records, expected 0 skipped, got 11

When I use the code in this PR or 41.0.0 to write out the same data, I can query the resulting parquet file successfully

$ datafusion-cli -c "SELECT xxx, time FROM '/tmp/foo.parquet' WHERE 1684850057953220316 <= time::bigint";
DataFusion CLI v27.0.0
+-------+---------------------+
| xxx | time                |
+-------+---------------------+
(there is data here)

Thus I am quite confident that this is the root cause and fix.

Thank you for the quick turnaround @tustvold

Test Program

fn main() {

    // read data from parquet and write it back
    let path = "/path/to/input/data.parquet";
    println!("reading from {path}");

    let file = File::open(path).unwrap();
    let builder = ParquetRecordBatchReaderBuilder::try_new(file).unwrap();
    let mut reader = builder.build().unwrap();


    // rewrite the data to /tmp/foo.parquet
    let path = "/tmp/foo.parquet";
    println!("writing to {path}");
    let file = File::create(path).unwrap();
    let mut writer = ArrowWriter::try_new(file, reader.schema(), None).unwrap();

    while let Some(record_batch) = reader.next().transpose().unwrap() {
        println!("Read {} records.", record_batch.num_rows());
        writer.write(&record_batch).expect("Writing batch");

    }

    // writer must be closed to write footer
    writer.close().unwrap();

}