feat(io): add ParquetFile with row-group statistics()

[yohplala]

Summary

Adds a new arro3.io.ParquetFile class that reads only the Parquet footer
(no data pages) and exposes per-row-group statistics as an Arrow RecordBatch,
implementing the design sketched in kylebarron#315.

The class wraps parquet::arrow::arrow_reader::ArrowReaderMetadata and uses
StatisticsConverter from the parquet crate to produce a three-column
RecordBatch (min, max, null_count), one row per row group. The min
and max columns take the Arrow data type of the source column; null_count
is UInt64.

from arro3.io import ParquetFile

pf = ParquetFile.open("data.parquet")
pf.num_rows, pf.num_row_groups, pf.num_columns

stats = pf.statistics("timestamp")  # arro3.core.RecordBatch
lo = stats.column("min")[0].as_py()     # first row group's min
hi = stats.column("max")[-1].as_py()    # last row group's max

Design notes

• Footer-only. ParquetFile.open calls ArrowReaderMetadata::load, which
  parses the Parquet footer and merges the embedded ARROW:schema metadata.
  No data pages are read, so construction is cheap.
• Class shape mirrors upstream PR Expand Parquet API: async, filtering, projection pushdown kylebarron/arro3#313. #[pyclass(module = "arro3.io", name = "ParquetFile", subclass, frozen)], #[classmethod] fn open(...),
  inner field meta: ArrowReaderMetadata, getter named schema_arrow
  (disambiguates from a future Parquet-schema accessor and matches pyarrow's
  pq.ParquetFile.schema_arrow). When PR Expand Parquet API: async, filtering, projection pushdown kylebarron/arro3#313 lands, this class should merge
  cleanly into its file layout.
• page_index=True is Optional, not Required. The upstream From<bool>
  impl maps true to PageIndexPolicy::Required, which errors if the page
  index is missing — a footgun for users opting in on files that lack one.
  We explicitly map true → Optional ("load if present, skip otherwise").
• null_count field nullability is !missing_null_counts_as_zero: when
  the default zero-fill path is taken the array is guaranteed to have no
  nulls, so the field is marked non-nullable; when the flag is False the
  field is nullable.
• Struct columns are not supported by the converter yet (upstream limit,
  Parquet StatisticsConverter does not work for struct columns apache/arrow-rs#7364). Documented on the method.
• Error handling flows through Arro3IoResult, matching the async path
  already in parquet.rs. No new .unwrap()s introduced.
• No dependency changes — parquet 58 already ships StatisticsConverter
  and ArrowReaderMetadata.

Changes

• arro3-io/src/parquet.rs — new PyParquetFile pyclass with open,
  num_rows, num_row_groups, num_columns, schema_arrow, __repr__,
  and statistics.
• arro3-io/src/lib.rs — register the class.
• arro3-io/python/arro3/io/_parquet.pyi — typed ParquetFile stub.
• arro3-io/python/arro3/io/_io.pyi — re-export ParquetFile (missing this
  would break type checkers and mkdocstrings).
• docs/api/io/parquet.md — docs entry.
• tests/io/test_parquet.py — 9 new tests covering accessors, int + float
  statistics (with pyarrow footer metadata as ground truth), nullable float
  column, first/last row-group bounds pattern, unknown column error,
  file-like object input, missing_null_counts_as_zero=False branch (field
  nullability), skip_arrow_metadata=True kwarg, and page_index=True on
  a file without a page index (pins the Optional-vs-Required choice).

Test plan

• uv run pytest tests/io/test_parquet.py — 12/12 pass (3 pre-existing + 9 new)
• uv run pytest tests/ — 170/170 pass
• cargo check -p arro3-io — clean
• cargo fmt -p arro3-io --check — clean
• Manual spot check: ParquetFile.open("data.parquet").statistics("timestamp")
  returns an arro3.core.RecordBatch with the expected schema, slicing
  first/last row group gives correct min/max, works with both a file
  path and a BytesIO file-like object.

Refs: kylebarron#315, apache/arrow-rs#7364