Refactor complex child column layout metadata

[JingsongLi]

Summary

Make ARRAY/MAP physical child layout explicit in the in-memory metadata used by bucket read/write paths. This keeps the on-disk format unchanged while removing order-based inference for nested complex child columns.

Changes

• Add ChildColumnRole plus length_physical_index to ChildColumnMeta so child columns know whether they are list elements, map keys, or map values, and which physical lengths column owns them.
• Refactor complex type expansion and writer traversal to locate nested children through explicit layout metadata instead of sibling index offsets.
• Refactor reader reassembly to rebuild ARRAY/MAP columns from explicit roles and length ownership rather than adjacent-child heuristics.
• Add a unit test covering the physical layout plan for a nested ARRAY<MAP<INT, ARRAY<UTF8>>> shape.

Testing

• cargo fmt --check
• git diff --check
• cargo test -p paimon-mosaic-core --lib
• cargo test -p paimon-mosaic-core --test array_type_test
• cargo test -p paimon-mosaic-core --test gen_fixtures

Notes

cargo test -p paimon-mosaic-core was also run. It passed through unit, ARRAY/MAP, encoding, format, and fixture tests, then failed in interop_read_test because /tmp/mosaic_interop/python_written.mosaic was not present; that test expects externally generated interop fixtures.

[QuakeWang]

LGTM

Non-blocking: the new test mainly verifies the metadata layout. It would be useful to add a real roundtrip case such as List<Map<Int32, List<Utf8>>>, covering the writer pending queue and reader reassembly together.