HIVE-29598 follow-up: virtualize clearValue + replace qtest

[konstantinb]

Follow-up to apache#6486 addressing two issues:

1. Architecture (responds to @soumyakanti3578's review comment): the
   clearVectorValue private static instanceof-chain dispatch is replaced
   by a virtual ColumnVector.clearValue(int) method with a template-method
   clearSlotValue(int) hook. clearValue is final and centralizes the
   isNull[] / noNulls bookkeeping; subclasses override only clearSlotValue,
   so the contract cannot drift across future call sites or new subclasses.
   The operator's call sites collapse to one-liners. Additive to storage-api,
   no breaking changes.

2. Test rigor: the original vector_outer_join7.q passes whether the
   fix is applied or not — it does not catch HIVE-29598. Confirmed by
   direct execution: pre-PR operator (97570b9cc2) + the author's
   original .q from 6b31424d46 produces output identical to the fixed
   operator. The CBO simplifies the plan (LEFT JOIN → ANTI JOIN, CASE WHEN
   → constant 'new') so aggressively that the null-padded columns are
   never read downstream.

   Replaced the .q with the actual repro from the HIVE-29598 description.
   New qtest:

   • Plan shows VectorMapJoinOuterMultiKeyOperator with
     smallTableValueMapping: 4:boolean — the slot reuse the JIRA predicts.
   • Output with fix: zero rows.
   • Output without fix: C 3 0 1 and D 3 0 1 leak through — exactly
     the rows the JIRA predicts.

Test plan

• mvn test -pl storage-api — 26 tests pass (4 new JUnit 5 files for
  Long/Double/Decimal/IntervalDayTime + 2 extensions on existing JUnit 4
  Bytes/Timestamp).
• mvn test -pl ql -Dtest=TestVectorMapJoinOuterGenerateResultOperator
  — 9 cases pass (3 individual covering Long-tracking + Void-baseline, plus
  6 parameterized over every modified ColumnVector subclass). Replaces the
  original PR's reflection-on-private-static test.
• vector_outer_join7.q passes with fix; FAILS with exactly the 2-row
  diff under pre-PR operator. Both .q.outs captured for comparison.

Out of scope

The three isNull[i] = true writes in FULL OUTER paths (overflowBatch
writes at the original lines 771/880/1014) are deliberately unchanged. A
FULL OUTER repro attempt confirmed those paths do fire
(VectorMapJoinFullOuterLongOperator), but the FULL OUTER planner inserts
IfExprCondExprNull wrappers that respect isNull[] and recover from the
stale-value bug class. Without a working wrong-results repro, expanding
the fix there is speculative.

---

Drafted with Claude Code; verified and posted by @konstantinb.

[ryukobayashi]

Thanks. I accept your proposal.