ParquetRecordBatchStreamBuilder::new() panics instead of erroring out when opening a corrupted file

[mmaitre314]

When opening a corrupted Parquet file where one of the row groups is missing a column, ParquetRecordBatchStreamBuilder::new() panics instead of returning an error:

 A panic occurred at (...)\parquet-50.0.0\src\file\metadata.rs:352: assertion `left == right` failed
  left: 2
 right: 1

This is due to the RowGroupMetaData::from_thrift() validating column counts using the assert_eq!() macro:

arrow-rs/parquet/src/file/metadata.rs

Line 352 in 639e81e

assert_eq!(schema_descr.num_columns(), rg.columns.len());

Could the check be replaced by an if statement returning an error? If that sounds OK, I can prepare a PR for review.

As workaround, it is possible to unwind the panic and handle it almost like a regular error. Unwinding async code is not straightforward though and it would be simpler to allow consistent error handling.

    let async_panic_handler = AssertUnwindSafe(async { ParquetRecordBatchStreamBuilder::new(blob_reader).await }).catch_unwind().await;
    if async_panic_handler.is_err() {
        info!("Skipping Blob (invalid metadata)");
        return Ok(())
    }
    let parquet_builder = async_panic_handler.unwrap()?;

[tustvold]

Changing to an error seems straightforward enough, but I will point out the actual thrift decoding logic can also panic.

FWIW if using tokio, it already intercepts and handles panics for you.

#3279

[tustvold]

label_issue.py automatically added labels {'parquet'} from #5362