[format][python] Coalesce map blob metadata reads

[XiaoHongbo-Hope]

Purpose

Coalesce adjacent MAP<X, BLOB> indexes in Python and Java. Java also buffers consecutive keys with bounded read-ahead, avoiding per-key reads without touching BLOB values.

Detect native commit APIs so older Rust runtimes retain Python fallback.

Tests

• Commit/BLOB tests: 242 passed with the development Rust runtime; 204 passed and 38 skipped with Rust 0.3.0.
• 170 Python BLOB tests and 44 Java BLOB tests passed; Flake8 and full Maven checks passed.
• Range-tracking tests cover descriptor correctness, null/empty values, buffer boundaries, and no value-data reads.
• Real OSS sample with 17 entries: Java metadata Range GETs 21 → 4, unchanged 854 bytes and descriptor checksum; median 0.931s → 0.217s over three runs per version. Measured with a direct Range-GET stream, excluding catalog/scan time.

[Akash3121]

Nit: avoid coupling this regression to the readers total req count
This exact count also includes unrelated file-footer and row-index reads, so a future optimization or refactor in those readers could break this test even if the map metadata remains correctly coalesced. Could we assert the relevant map index/key ranges directly, or use an upper bound here as in the large-metadata case? That would preserve the optimization guarantee without coupling the test to the complete reader request sequence.

[JingsongLi]

+1