[core] Align primary-key vector source metadata with Java

[JunRuiLee]

Summary

• parse Java's current primary-key vector source metadata frame with dataLevel
• filter current ANN segments by exact level source coverage, matching Java PkVectorBucketIndexState
• update PK-vector source-meta fixtures and tests

Motivation

This follows the Java-side PK-vector source metadata changes in apache/paimon#8672.

Java now writes dataLevel into GlobalIndexMeta.sourceMeta between the frame version and source-file count. Without this update, Rust reads dataLevel as the source-file count and cannot correctly consume Java-written PK-vector metadata.

Java also selects current ANN payloads by exact level coverage: active compacted files are grouped by data level, each segment's sourceMeta.dataLevel() must match one level, and the segment's source files must exactly equal that level's active files. Mirroring that rule keeps Rust from searching stale or partially-overlapping ANN segments after compaction rewrites files.

Tests

• cargo fmt --check
• cargo test -p paimon pk_vector_source --lib
• cargo test -p paimon pk_vector_scan --lib
• cargo test -p paimon vindex::pkvector::bucket --lib
• cargo test -p paimon vindex::pkvector::ann --lib
• cargo test -p paimon pk_vector_orchestrator --lib
• cargo test -p paimon --test pk_vector_baseline_test
• cargo clippy -p paimon --all-targets -- -D warnings

[JingsongLi]

+1