[spark] Support merge-on-read for postpone bucket tables

[Zouxxyy]

Purpose

Postpone bucket tables buffer batch-written records in the postpone bucket, so these records are not visible to ordinary queries until postpone compaction finishes.

This change introduces opt-in merge-on-read support through spark.paimon.postpone.merge-on-read=true.

• Add an engine-neutral Core plan and read API for planning real-bucket splits and physical postpone-file tasks from one snapshot.
• Preserve postpone-file replay order with deterministic sequence ranges, route postpone records to their target buckets, and reuse Paimon's existing merge engine for each bucket.
• Add a Spark DSv2 scan and physical execution path which clusters real split metadata and postpone records by partition and target bucket before delegating the merge to Core.
• Fall back to the ordinary Paimon scan when no postpone files are present, keeping the existing pushdowns and execution path.
• Share bucket assignment and postpone-file ordering between merge-on-read and the Spark postpone compact procedure.

The option is disabled by default, so existing delayed-visibility behavior is unchanged.

Tests

• Core targeted tests: PostponeUtilsTest and TableScanTest#testPostponeMergeReadBuilderAndPushDown — 11 tests passed.
• Spark PostponeBucketTableTest — 14 tests passed.
• Spark PrimaryKeyVectorSearchTest — 12 tests passed.
• Compiled the Spark 3.2, 3.5, and 4.1 modules.
• Ran Spotless, regenerated the Core configuration documentation, and verified git diff --check.

[JingsongLi]

PR generates a global sequence in PostponeUtils based on creationTime → fileName → line order within the file. However, the actual behavior of Postpone should be to guarantee the file order for a given (table, partition, writeId); when different writers enter the same bucket writer in parallel, the final order depends on the arrival order of the multi-input channels in Flink.

[JingsongLi]

create() first checks a snapshot to determine whether to follow a special execution path, while plan() re-parses the latest snapshot; they should use the same snapshot.