[python] Add opt-in late materialization using Parquet OffsetIndex

[XiaoHongbo-Hope]

What

• Add opt-in late materialization for row-ID-based Parquet reads. Once Paimon identifies the requested rows, use OffsetIndex to fetch only their data pages and let PyArrow decode them.
• Support flat columns and standard two-binary-field VARIANT (value, metadata). A single row is a valid one-row range. Disjoint ranges, full-row-group reads, missing indexes, unsupported schemas, and selections that would read too many bytes use the ordinary reader.
• Reuse parquet.filter.columnindex.enabled, disabled by default. This does not add ColumnIndex predicate filtering or change the table format.

I/O trade-off

The optimization trades more small range requests for fewer transferred bytes. Read-only single-row tests returned identical results and showed lower file-read and point-lookup latency, but issued more GET requests. The extra requests read page indexes and may outweigh byte savings on high-latency storage or less selective reads. This is why the path remains opt-in and falls back when it is unlikely to help. These measurements are not a full application or training-sample benchmark.

Validation

• Related read tests: 108 passed, 18 subtests passed (PyArrow 18); page-index tests: 61 passed (PyArrow 19).
• Python lint, Python 3.6 syntax check, and git diff --check passed.

This is a draft because the remote I/O behavior and fallback thresholds deserve community review before enabling it by default.