[python] Batch BTree point reads by data block

[XiaoHongbo-Hope]

Purpose

Avoid reading the same BTree SST data block once per literal for batched IN predicates.

Changes

• Group serialized point keys by their selected SST data block.
• Read each selected block once and resolve its keys in memory.
• Deduplicate repeated literals without retaining a cross-query cache.

Tests

• Added a regression test proving four keys in one data block issue one data-block read.
• 102 passed across global-index, BTree thread-safety, and vector-filter tests.
• Flake8 and git diff --check pass.