[core] Optimize tag reads during snapshot expiration

[JingsongLi]

Purpose

Snapshot expiration previously built a full data-file index for every referenced tag. On large tables, concurrent tag reads retained memory proportional to the number of tags multiplied by all table files and could cause OOM.

This change:

• plans deletion candidates in bounded expiration batches;
• prunes tag manifests by candidate bucket ranges;
• pushes partition and bucket filtering into Avro decoding before nested file metadata;
• filters file names on a reusable projected entry before materializing ExpireFileEntry;
• keeps only matching tag entries while preserving changelog-decoupled cleanup behavior.

The filename filter is streaming rather than storage-index based; manifests still need to be scanned within matching partitions and buckets.

Tests

• Targeted manifest streaming/filter-pushdown and tag expiration regression tests
• ExpireSnapshotsTest and FileDeletionTest
• mvn -pl paimon-core -DskipTests spotless:check

[leaves12138]

LGTM. The candidate-scoped tag reads reduce retained state while preserving tag protection and changelog-decoupled cleanup behavior. Repeated scanning across batches is an acceptable trade-off for reducing memory pressure.

Validated with Java 8: 161 existing test cases across ManifestFileTest, ExpireSnapshotsTest, FileDeletionTest, ChangelogExpireTest, and TagTest, plus a local cross-batch tag-retention regression (162 passed in total). No blocking issues found.