fix: correct query result memory estimate

[chrispader]

Summary

• correct unordered_map row node memory accounting to add per-node padding instead of multiplying by it
• include row key strings, large text values, owning blob buffers, and metadata map storage in query result external-memory pressure
• move query result containers into the HybridObject to avoid extra peak memory copies
• avoid building full query result objects during batch execution when only rowsAffected is needed

Validation

• nr lint-cpp
• nr sqlite typecheck
• git diff --check

Fixes the issue reported in #261 (comment)