Are there improvements in how memory_limit works? I often had DuckDB get OOM killed because it went beyond its limit. It's definitely one of the reasons why I usually have an AI tune the environment for my datasets.