scan: init external rel table scan state before first scan()

[adsharma]

Fixes the regression introduced by the pg_client extension moving rel table registration to RelGroupCatalogEntry + ForeignRelTable: MATCH (a:node_person)-[k:rel_knows]->(b:node_person) was crashing with tableFunc: sharedState is null on the first scan.

Root cause

ScanRelTable::initLocalStateInternal creates the per-thread scan state but only calls table->initScanState in fetchNextBoundNodeBatch() and after pulling a new child tuple. The first scan() in getNextTuplesInternal therefore runs against an uninitialized scan state.

The native RelTable::initScanState reads from the bound-node nodeIDVector to pick a node group, so it must run after a child tuple is in flight. External rel table backends (ForeignRelTable, ArrowRelTable, IceDiskRelTable) only use initScanState to set up scan-function / shared / local state and source/nodeGroupIdx — they don't need a bound node — but the first scan() still fires before any re-init call site, so we'd dereference uninitialized state and crash.

Fix

Detect "external" via typeid(*tableInfo.table) != typeid(RelTable) and call table->initScanState once after creating the per-thread scan state, so the first scan() is well-defined. This scales to any future external backend (XxxRelTable : RelTable) without re-enumerating.

Native RelTable behavior is unchanged: first scan() still returns false (source == NONE), then fetchNextBoundNodeBatch re-inits with the child tuple in flight.

Companion change

The pg_client extension submodule was updated to restore RelGroupCatalogEntry registration for rel_* tables (so the planner has a real rel entry to anchor MATCH on) and to add a mutex around the morsel currentOffset in tableFunc (so parallel HashJoinBuild workers atomically claim rows). That lives in the extensions repo on the pg_client branch.

Tests

Verified with the pg_client extension's Python test suite (13/13 passing, including new test_06b_match_rel_table / test_06c_match_rel_count_parallel) and the e2e runner (new MatchOnRelTable case). The push-down optimizer (ForeignJoinPushDownOptimizer) correctly lowers MATCH ... RETURN count(*) to a single COUNT_REL_TABLE operator on rel_knows.