[python] Support reading shared-shredding maps

[XiaoHongbo-Hope]

Purpose

PyPaimon currently exposes the physical struct when reading a MAP<STRING, T> stored with the shared-shredding layout. Restore the complete logical map instead.

Changes

• parse shared-shredding field metadata with none, lz4, and zstd dictionary compression
• assemble fixed physical columns and overflow entries into the logical Arrow MAP
• preserve null and empty maps and arbitrary supported value types
• recover Paimon field metadata from the ORC ARROW:schema entry across supported PyArrow versions
• restore ORC physical representations of TIME, TIMESTAMP, and TIMESTAMP_LTZ, including values nested in rows, arrays, and maps
• resolve shared-shredding metadata through source paths so nested-projection output aliases remain transparent

Selected-key projection and shared-shredding writes are intentionally out of scope.

Tests

53 related reader tests passed locally on Python 3.11 / PyArrow 19. The seven shared-shredding tests also passed on Python 3.6 / PyArrow 6; the existing nested VARIANT reader suite passed there as well. Java-written ORC files using all three metadata compressions and TIME values were verified with both PyArrow versions. PyArrow-written ORC reader regressions cover TIME, TIMESTAMP precision, and nested TIMESTAMP_LTZ values. The projection regression covers a nested a.b output colliding with a complete shared-shredding a_b column. A mixed-language E2E now has Java write Parquet and ORC shared-shredding MAPs across two files, and Python verify fixed/overflow keys, null values, empty maps, null maps, and per-file dictionaries. Flake8, Spotless, shell syntax, and git diff --check passed.

[JingsongLi]

Please add some E2E test cases.

[XiaoHongbo-Hope]

Please add some E2E test cases.

Added

[JingsongLi]

+1