[Flink] Add configurable initial scan mode for streaming compaction

[lilei1128]

Purpose

When a dedicated streaming compaction job starts for a table with existing
snapshots but no COMPACT snapshot, it scans from the earliest available
snapshot to catch up with the historical backlog. For tables with many
snapshots, this catch-up can take a long time, while snapshot expiration may
remove snapshots that are still being scanned. The compaction job can then
fail and repeatedly restart.

Add compaction.initial-scan-mode to allow the initial compaction scan to use
the latest snapshot as an ALL baseline when historical changelog replay is not
required. The default remains earliest to preserve existing behavior and
changelog semantics. When a COMPACT snapshot exists, compaction continues from
the snapshot after the latest COMPACT snapshot.

Tests

Add tests covering:

• default earliest behavior;
• latest baseline behavior without a COMPACT snapshot;
• latest baseline followed by incremental delta scans;
• multiple partitions and buckets in the latest baseline;
• existing COMPACT snapshot behavior.

[JingsongLi]

it is for ContinuousCompactorStartingScanner, I think the name should be related to continuous.

[lilei1128]

continuous-compaction.initial-scan-mode is more appropriate name.