fix: bad downcast of ArrowResultCollector in plan mapper re-enables skipped arrow/CSR tests (#881)

[adsharma]

Fixes #881.

Root cause

ArrowResultCollector and the regular ResultCollector both declare
static constexpr PhysicalOperatorType type_ = PhysicalOperatorType::RESULT_COLLECTOR.
In PlanMapper::getPhysicalPlan() the plan-root marker introduced by #870/#877 work was applied via an operator-type check:

if (root->getOperatorType() == PhysicalOperatorType::RESULT_COLLECTOR) {
    root->ptrCast<ResultCollector>()->setResultExposedToClient();
}

For every arrow query (resultType=ARROW), the newly created ArrowResultCollector matched that check and was then downcast to ResultCollector, which it is not:

• With RUNTIME_CHECKS (or any !NDEBUG build), dynamic_cast_checked asserts and throws. Every queryAsArrow() execution failed and returned an error MaterializedQueryResult whose iterator is null. Tests calling getNextArrowChunk() without checking isSuccess() then SIGSEGV in FactorizedTableIterator::hasNext — the Fatal signal 11 in the linux minimal-test job. The remaining tests fail on dynamic_cast<ArrowQueryResult*>(result.get()) == nullptr and entry.relCsrResults[0] == nullptr (PROJECT_GRAPH materialization also runs through queryAsArrow).
• With NDEBUG (plain release), the checked cast is a reinterpret_cast and setResultExposedToClient() writes a byte at ResultCollector::internalResultTable's offset (192), which in an ArrowResultCollector object is localState.batchIndex — UB on every arrow query.

The minimal-test job builds with -DENABLE_RUNTIME_CHECKS=1, which made the failure deterministic across all 10 arrow/CSR tests — while local (non-RUNTIME_CHECKS) runs looked healthy, which is why the issue looked like a timing-dependent race.

Fix

Mark the client-facing result table on the freshly created regular ResultCollector inside the non-arrow branch of the if/else, instead of matching on the shared operator-type enum afterwards.

Also re-enables the ten tests skipped in 520abda.

Note on the earlier investigation

The "corrupted task clone" evidence in the issue (this == sharedState, garbage batch indices, CSR metadata present for CSR-free queries) was an artefact of the temporary LBUG_ARROW_DEBUG tracing: its fprintf macro appended the thread-id string after the caller's varargs while the format consumed %s first, shifting every printed value by one slot. The deterministic "reproduction with instrumentation" was the tracing itself crashing in strlen() on an integer consumed as %s. With corrected tracing, all collector state is healthy up to the crash.

Validation

• Full api_test (304 tests) passes with -DCMAKE_BUILD_TYPE=RelWithDebInfo -DENABLE_RUNTIME_CHECKS=1 (the minimal-test configuration), plain RelWithDebInfo, and Debug builds.
• TSAN run of the arrow/CSR suite is clean.