[python] Fix Parquet VARIANT reads with multi-chunk nested data

[XiaoHongbo-Hope]

Purpose

PyArrow can raise Nested data conversions not implemented for chunked array outputs when projected VARIANT data is decoded into multiple chunks, including within one row group or inside ROW, ARRAY, and MAP columns.

Use bounded ParquetFile.iter_batches() reads while preserving projection, predicates, filter-only columns, and row-group pruning. Exact post-read projection avoids dotted top-level name collisions, and nested shredded VARIANT values are assembled recursively.

Tests

• Single-row-group multi-chunk reproducers for top-level and nested VARIANT.
• Dotted top-level and nested-path projection regressions.
• Plain and shredded VARIANT in ROW, ARRAY, and MAP, including NULL and overflow values.
• PyArrow 19: 169 passed, 3 subtests passed.
• PyArrow 6 targeted compatibility: 13 passed, 3 skipped.