Microbenchmark doesn't support BHJ

[FelixYBW]

Description

Like TPCDS Q67, first stage.

It reports segment fault. Call stack:

#0  0x00007ffff1425b9b in JavaVM_::GetEnv (version=65544, penv=0x7fffffff7df8, this=0x0) at /usr/lib/jvm/java-17-openjdk/include/jni.h:1946
#1  gluten::JniHashTableContext::callJavaGet (this=0x7ffff7d74ea0 <gluten::JniHashTableContext::getInstance()::instance>, id=...)
    at /incubator-gluten/cpp/velox/jni/JniHashTable.cc:50
#2  0x00007ffff159ab58 in gluten::SubstraitToVeloxPlanConverter::toVeloxPlan (this=0x7fffffff9ec0, sJoin=...)
    at /incubator-gluten/cpp/velox/substrait/SubstraitToVeloxPlan.cc:467
#3  0x00007ffff159a2c9 in gluten::SubstraitToVeloxPlanConverter::toVeloxPlan (this=this@entry=0x7fffffff9ec0, rel=...)
    at /incubator-gluten/cpp/velox/substrait/SubstraitToVeloxPlan.cc:1686
#4  0x00007ffff159ce35 in gluten::SubstraitToVeloxPlanConverter::convertSingleInput<substrait::ProjectRel> (rel=..., this=0x7fffffff9ec0)
    at /incubator-gluten/cpp/velox/substrait/SubstraitToVeloxPlan.h:273

Gluten version

None

[malinjawi]

#12592 fixes the segfault, but the benchmark still cannot replay a BHJ: the build side input iterator is Iterator.empty on the offload path (VeloxSerializedBroadcastRDD.scala:47, VeloxBroadcastBuildSideRDD.scala:50), so the dumped data_*.parquet for that index has zero rows.

Proposed approach, all native — no new JNI, no Scala changes:

1. At VeloxRuntime::parsePlan time the table is already in HashTableCache: getIterators populates it (GlutenWholeStageColumnarRDD.scala:77) before genFirstStageIterator reaches the native kernel.
2. So VeloxWholeStageDumper can walk the parsed plan for JoinRel.hashtableid, look each up in HashTableCache::instance(), serialize with the existing HashTableSerializer, and write hashtable_{stageId}_{partitionId}_{vId}_{id}.bin next to the plan. ignoreNullKeys falls out of the HashTable<true>/<false> dynamic_cast; joinHasNullKeys is entry->hasNullKeys.
3. GenericBenchmark gains --hash_table, calls the existing gluten::deserializeHashTable(...), and registers via HashTableCache::instance()->add(...) — mirroring deserializeHashTableDirect.
4. getJoin() consults HashTableCache when there is no JVM, so the converter sets useHashTableCache=true and the probe runs against the real table.

HashTableCache::add sets builderTaskId = "external_gluten" and buildComplete = true, so HashBuild::setupCachedHashTable short-circuits to noMoreInput() and never reads the build side — the empty parquet stays harmless. An externally added entry also skips the queryCtx release callback in HashTableCache::get, so it survives across benchmark iterations.

Happy to implement if this looks reasonable.