MR: Keep computed block locations in IcebergSplit.getLocations

[thswlsqls]

Closes #18032

Summary

• IcebergSplit declares transient String[] locations and checks locations == null, so the value is meant to be computed once and cached. The else branch overwrites it with ANYWHERE on every later call, so two calls on an unchanged split return different values — Hadoop's InputSplit.getLocations() contract expects them to match.
• The else branch came in with #1582, which fixed an NPE when getLocations() runs on a worker node whose deserialized split has a null conf. The original implementation, #1192, had if (locations == null) alone. This change restores that shape and keeps the guard by folding the null check into localityPreferred, so a null conf still resolves to ANYWHERE and never calls Util.blockLocations.
• No production failure is claimed here: plain MapReduce submission calls getLocations() once per split, so the case rests on the contract and on the caching the class already declares.

Testing done

• Extended TestIcebergInputFormats#testLocality to call getLocations() twice on the same split and assert both calls return the computed locations. It fails before the fix — all 6 parameterized variants return ["*"] on the second call — and passes after.
• ./gradlew :iceberg-mr:check — 158 tests, 0 failures.

---

AI Disclosure

• Model: Claude Opus 5
• Platform/Tool: Claude Code
• Human Oversight: [unknown - human to fill in]
• Prompt Summary: Fix IcebergSplit.getLocations() so a cache hit returns the previously computed block locations instead of overwriting them with ANYWHERE, keep the null-conf NPE guard added in MR: Fix NPE when InputSplit.getLocations is called on mappers #1582, and extend testLocality to cover repeated calls.