[FEA] `PythonScan` support row-limits in multi-rank GPU execution

## Summary

cudf-polars currently rejects pushed-down row limits on `PythonScan` sources, falling back to the CPU engine. The underlying issue is that a single global row limit cannot be enforced independently on each rank of a distributed query. This issue tracks adding proper distributed support so that `head`, `limit`, and `tail` (surfaced as `n_rows`) work correctly across ranks instead of being rejected.