Sub-plan reuse scoped by digest barriers

Applies to: every query with a shared source or self-join (multi-view/self-join queries)

Installing the native planner used to disable Flink's sub-plan reuse outright — the safe-but-blunt way to keep any Arrow batch from fanning out to two consumers (the hand-off is zero-copy; the consumer closes the buffers). That also un-shared the rowwise prefix ahead of the island, so every multi-view/self-join query generated and converted its source stream once per branch: the profiling round measured an exactly-2x Row→RowData conversion tax on q3/q4/q5/q7/q8/q9/q20.

The fix keeps reuse enabled and adds a per-instance term to every native rel's digest (emitted only at the digest explain level, so EXPLAIN output is unchanged). Flink's post-optimize reuse pass merges by digest, so it can now merge the shared rowwise prefix under the islands exactly as stock Flink does — but it can never merge a columnar subtree, since each native instance's digest is unique.

Measured on the generator profile loop: q3 +17%, q9 +9%, q20 +6%, with the conversion cost per iteration restored to parity with stock Flink's.