Answer COPY TO with RETURN_STATS as the rows, sizes and column statistics of each Parquet file

[tamnd]

Part of #1446.

COPY ... TO a Parquet file now takes RETURN_STATS. The copy answers one row per file with filename, count, file_size_bytes, footer_size_bytes, column_statistics, partition_keys and extra_info, with the same types as the pin.

column_statistics maps each column, named in double quotes with inner quotes doubled, to its column_size_bytes, null_count, num_values and, when the column has values, its bounds with min_is_exact and max_is_exact. The keys are sorted, as the pin sorts them. A bool reports 1 and 0, a blob reports its bytes in hex, and a string or blob longer than 256 bytes is cut short and marked as not exact, with the upper bound raised so it stays above every value. partition_keys holds the partition values as text in the order of PARTITION_BY and is NULL for a copy that is not partitioned. extra_info holds the row group count.

A float column also reports has_nan and nan_count, counted over every row group. This follows the corpus, which comes from a DuckDB newer than the pin. v2.0.0-dev84237 has no nan_count and only keeps has_nan from the last row group it wrote. A column with no values reports neither, as in the pin.

The Parquet writer keeps running totals for each column as it writes row groups, and hands back the footer size when it finishes. The sizes are the writer's own, so they differ from the pin's byte for byte, which the corpus allows for with a regex.