arrow

ARROW-10131: [C++][Dataset][Python] Lazily parse parquet metadata

ParquetFileFragment now constructs `Expression`s from its statistics lazily; the min/max expression for a column is materialized only when a predicate references that column.

Additional changes:
- ParquetFileFragment now simply stores a parquet::FileMetaData, which is loaded opportunistically anytime IO becomes unavoidable.
- In python, accessing any RowGroup or file-level properties of ParquetFileFragment will force load of metadata (so for example `ParquetFileFragment.num_row_groups` no longer yields -1 to indicate that metadata has not been loaded).
- RowGroupInfo has been removed from C++. A python-only replacement remains for compatibility but we might want to remove that as well.
- Reduced path manipulation in ParquetDatasetFactory, including avoidance of validation of ColumnChunk paths by default.
- ParquetScanTaskIterator has been removed
- Added FileMetadata::Subset, which returns a FIleMetaData wrapping only a subset of row groups.
- Added native equality comparison between FileMetaData, RowGroupMetaData, ColumnChunkMetaData, Statistics (ARROW-4970)

Closes #8507 from bkietz/10131-Lazily-parse-parquet-meta

Authored-by: Benjamin Kietzman <bengilgit@gmail.com>
Signed-off-by: Joris Van den Bossche <jorisvandenbossche@gmail.com>