Add end-to-end integration test for multi-region workflow execution #6133

Task Summary

Cross-region data delivery goes through input-port materialization reader threads (which replaced the cache-source operators in #3425). This path has no end-to-end coverage — it is exercised only by scheduling unit tests that count regions on a physical plan, so a regression in the materialization read/write path could pass CI unnoticed.

This matters right now because the now-dead cache-source plumbing is about to be removed (#6125 / #6126). The multi-region path — exactly what the removed cache sources used to serve — should get end-to-end coverage before that refactor.

Add an integration test that drives a big "X"-shaped workflow across multiple regions:

csvLeft  ─┬─▶ join.build (port 0) ─┐
          │                         join ─▶ pythonSort ─▶ (terminal)
csvRight ─┼─▶ join.probe (port 1) ─┘
          │
csvLeft  ─┼─▶ union ───────────────────────────────────▶ (terminal)
csvRight ─┘   (two links fan into union's single port)

The hash join's probe-depends-on-build ordering forces the build input to be materialized, cutting the plan into ≥2 regions whose boundary is crossed by the reader-thread path. A Python sort UDF makes it a real end-to-end run with worker subprocesses.

Assertions:

Notes:

• Python-tagged (@IntegrationTest) so it runs in the amber-integration CI job (which provisions Python deps); the lighter amber job excludes it.
• Should land before refactor(engine): remove dead CacheSourceOpExec and OpExecSource plumbing #6126, which is then rebased on top so its CI re-runs this test against the removal.

Task Type

• Test / Coverage