use a single row_count column during predicate pruning instead of one per column

[adriangb]

Closes #13836

[adriangb]

@alamb this seems to be a simple enough change that is almost self contained: the only breaking change that I see would be in PredicateRewriter which we recently introduced in #12850. I can't be sure of course but I'd guess we (Pydantic) are the only ones using this, we'd be happy to absorb the breaking change. The alternative would be to add an option to PredicateRewriter to control this behavior, which requires more code, etc. Up to you 😄.

[adriangb]

I want to point out that this works because of how the Recordbatch is generated:

datafusion/datafusion/physical-optimizer/src/pruning.rs

Lines 937 to 965 in 20544bc

	for (column, statistics_type, stat_field) in required_columns.iter() {
	let column = Column::from_name(column.name());
	let data_type = stat_field.data_type();
	let num_containers = statistics.num_containers();
	let array = match statistics_type {
	StatisticsType::Min => statistics.min_values(&column),
	StatisticsType::Max => statistics.max_values(&column),
	StatisticsType::NullCount => statistics.null_counts(&column),
	StatisticsType::RowCount => statistics.row_counts(&column),
	};
	let array = array.unwrap_or_else(|| new_null_array(data_type, num_containers));
	if num_containers != array.len() {
	return internal_err!(
	"mismatched statistics length. Expected {}, got {}",
	num_containers,
	array.len()
	);
	}
	// cast statistics array to required data type (e.g. parquet
	// provides timestamp statistics as "Int64")
	let array = arrow::compute::cast(&array, data_type)?;
	fields.push(stat_field.clone());
	arrays.push(array);
	}

Since it's generated based on the columns tracked by RequiredColumns we can just rename it internally with no consequences.

This should also save some work in creating the array, make scanning the record batch faster, etc.