[core] Reduce global index planning IO

[XiaoHongbo-Hope]

Purpose

Align Rust global-index planning with Java and PyPaimon to avoid unnecessary object-store IO and unrelated index candidates.

Changes

• prune sorted global-index manifest entries with the scan partition filter before evaluating index shards, matching Java and PyPaimon
• use the file size already stored in IndexFileMeta, falling back to metadata lookup for legacy or invalid sizes, matching GlobalIndexIOMeta in Java and PyPaimon
• keep deletion-vector selection and reader cache behavior unchanged

Performance

On a partitioned data-evolution table with an equality lookup and a partition predicate:

• native FAST planning: 8.18s -> 2.93s
• retained index entries: 762 -> 176
• indexed candidate ranges: 4 -> 1
• peak RSS in a plan-plus-explain benchmark: about 2.45 GiB -> 586 MiB

The result split and row range were unchanged.

Tests

• cargo test -p paimon global_index_scanner --lib (46 passed)
• cargo test -p paimon table_scan::tests --lib (84 passed)
• cargo fmt --all -- --check
• git diff --check

[QuakeWang]

LGTM

[JingsongLi]

+1