[FEA] Polars engine CPU fallback for GPU out-of-memory errors

[beckernick]

Today, users of the Polars GPU engine don't experience graceful performance degradation for GPU out-of-memory errors like users of cuDF pandas do -- instead, the query will fail. This is a pain point that's come up consistently in feedback from early users.

In parallel to investigating UVM, we should also investigate the feasibility of the current inspector-executor design supporting different types of CPU fallback when the GPU engine hits an out-of-memory error. E.g., we could rewind and execute the full query from scratch with the CPU engine or we could rewind only a portion of the query (which might require subgraph based hybrid execution).