Key parquet column options by leaf path so nested columns get compressed - #8674

Fixes #8671

Summary

pq.ParquetWriter matches per-column compression / use_dictionary / column_encoding against the leaf column paths (train.list.element.input...), but both ParquetWriter._build_writer and ParquetDatasetWriter._write keyed them by top-level feature name. Nested list/struct columns never matched, so they fell back to pyarrow's defaults: uncompressed and no dictionary encoding.

This moves the options into one get_parquet_column_options(features, schema) helper in arrow_writer.py that expands each column to its leaf paths, and uses it at both sites. Media columns (Image, Audio, ...) are structs too, so their "none"/PLAIN settings only held because unmatched leaves fell back to the defaults. The output for them is unchanged; the settings now actually match those leaves.

Repro from the issue after the fix:

question                                             -> SNAPPY
train.list.element.input.list.element.list.element   -> SNAPPY
train.list.element.output.list.element.list.element  -> SNAPPY

Test

Added test_parquet_writer_compresses_nested_columns in tests/test_arrow_writer.py (flat, nested list/struct, Array2D, Image). It fails on main and passes with the fix. tests/test_arrow_writer.py, tests/io/test_parquet.py, tests/packaged_modules/test_parquet.py and tests/features/test_image.py pass; make quality is clean.