[flink] Lake-batch fallback should use KV snapshot instead of reading log from earliest

[fresh-borzoni]

Search before asking

• I searched in the issues and found nothing similar.

Description

Follow-up to #3296.

In batch mode on a lake-enabled PK table with no lake snapshot yet, the fallback reads every bucket's log from EARLIEST, ignoring existing Fluss KV snapshots. This is OOM-prone on large never-tiered tables.

Spark fixed the equivalent in #3317 with per-bucket dispatch (snapshot+tail where KV snapshot exists, log-only otherwise).

Flink can't 1:1 port because FlinkSourceSplitReader has no sort-merge on the Fluss-only path. We may reuse LakeSnapshotAndLogSplitScanner (already used by Flink's lake reader, already sort-merges per PK) for the fallback or do smth similar

Willingness to contribute

• I'm willing to submit a PR!