[core] Support full-text search for primary-key tables

[JingsongLi]

Summary

Add compaction-visible full-text indexing and search for primary-key tables. The index follows the primary-key table lifecycle, filters physical rows with deletion vectors, and incrementally consolidates immutable archives with the shared primary-key index LSM policy.

Changes

• Add primary-key full-text configuration, schema validation, deterministic analyzer fingerprints, and source metadata for ordered multi-file archives.
• Build bucket-local native full-text archives from level-1-or-higher compact output and maintain them asynchronously with shared fanout and stale-ratio compaction options.
• Support multi-source archive row-ID mapping, deletion-vector offsets, atomic publish/rollback, stale-source reclamation, and compatibility with existing single-source archives.
• Add snapshot-scoped full-text scan/read/search over physical data-file positions, deterministic RRF ranking, temporary-index fallback for uncovered eligible files, and hybrid vector/full-text fusion.
• Expose primary-key full-text search through Spark SQL and the Flink full_text_search procedure, with documentation for configuration, visibility, maintenance, and limitations.

Testing

• JDK 8: 21 focused paimon-core test classes (144 tests)
• Spark 3 / JDK 8: FullTextSearchTest and HybridSearchTest (13 tests)
• Spark 4 / JDK 17: FullTextSearchTest and HybridSearchTest (13 tests)
• Flink 1 / JDK 8: FullTextSearchProcedureITCase
• Flink 2 / JDK 11: FullTextSearchProcedureITCase
• JDK 8: clean paimon-core and paimon-full-text compile with Checkstyle, Spotless, and Maven Enforcer
• Native primary-key full-text contract test invoked; it runs when PAIMON_FTINDEX_JNI_LIB_PATH is available

Notes

• Primary-key full-text uses the fixed full-text implementation; no index.type option is introduced.
• Level-0 foreground files remain outside full-text coverage until compaction.
• The first release supports one character full-text column per primary-key table.
• Python API changes are intentionally out of scope.