[FEA] Support multi-partition `Filter` operations containing non-pointwise expressions

**Is your feature request related to a problem? Please describe.**
Motivated by https://github.com/rapidsai/cudf/pull/20075

We don't currently support non-pointwise filters for multiple partitions. E.g.:

```python
q = df.filter(pl.col("total_revenue") == pl.col("total_revenue").max())
```

**Describe the solution you'd like**
We can probably re-use the expression-decomposition logic we already use for `Select` operations.

**Describe alternatives you've considered**
Rely on single-partition fallback.