Arrow: Fix NPE reading a constant/null column in vectorized reader

[thswlsqls]

Closes #10275

Summary

• When a column is added with ALTER TABLE ... ADD COLUMN after data files were written, the arrow vectorized reader produces a constant VectorHolder whose vector is null.
• GenericArrowVectorAccessorFactory.getPlainVectorAccessor called vector.getClass() on that null vector, throwing a confusing NullPointerException.
• Guard the null vector so the message becomes Unsupported vector: null, matching the null-safe pattern already used by the default branch in the same class (line 178).
• Revives the stale fix from #10275 - fix NullPointerException #10284, closed by the stale bot (not rejected by design); adds the reproduction test the reviewer requested, with no Spark dependency.
• Fully reading added constant/null columns in the vectorized path is out of scope here and tracked as follow-up on NullPointerException when using VectorizedArrowReader to read a null column #10275.

Testing done

• Added TestArrowReader#testReadAddedColumnFailsWithClearMessage: writes one row, adds a column via updateSchema().addColumn, scans with VectorizedTableScanIterable, and asserts UnsupportedOperationException with message Unsupported vector: null.
• Verified the test fails on main (NPE on vector.getClass()) and passes with the fix.
• ./gradlew :iceberg-arrow:check — passed (30 tests, JDK 21).

---

AI Disclosure

• Model: Claude Opus 4.8
• Platform/Tool: Claude Code

[singhpk234]

Is it then just throwing UnsupportException instead of NPE ? we should update the pr description accordingly
because the fix is not making the reads successful just wraps into another exception, is the workaround then to disable vectorized reads ?

[thswlsqls]

You're right — this doesn't make the read succeed; it replaces the confusing NPE (getPlainVectorAccessor calling getClass() on a null vector) with a clear UnsupportedOperationException: Unsupported vector: null. Fully reading an added constant/null column in the vectorized path is still out of scope here and tracked on #10275; the workaround today is to disable vectorized reads. I'll update the description so it reads as "clearer failure, not a successful read."