Add UNION ALL support for gpu_processing (legacy Sirius)

[ducndh]

gpu_processing queries with UNION ALL fall back to CPU

Implementation:

• Added GPUPhysicalUnion operator (src/legacy/operator/gpu_physical_union.cpp) following the same pipeline construction pattern as DuckDB's PhysicalUnion — creates a union pipeline that shares the same sink as the current pipeline, builds each child into its respective pipeline, and respects order-dependent sinks.
• Added gpu_plan_set_operation.cpp to map LogicalSetOperation → GPUPhysicalUnion in the plan generator. Only UNION ALL is supported; EXCEPT/INTERSECT throw NotImplementedException for now but you can try to build them
• Wired up CreatePlan(LogicalSetOperation&) in GPUPhysicalPlanGenerator to route LOGICAL_UNION through the new operator.
• Refactored GPUResultCollection from raw DataChunk* array with manual new[]/delete[] to vector<unique_ptr>, fixing a potential issue when multiple pipelines append results concurrently.
• Fixed pipeline dependency scheduling in gpu_executor.cpp to check dependencies across all pipelines in a meta-pipeline, not just the base pipeline — without this, UNION ALL pipelines could be scheduled before their build-side dependencies completed.