Smarter default for `split_out` in groupby aggregations

Currently, the default for `split_out` is based on the number of partitions and groupby columns:

https://github.com/dask-contrib/dask-expr/blob/fdeee4375df55499462bc35af50d3d216d5e256c/dask_expr/_groupby.py#L81-L82

Ideally, this would take (an estimate of) the number of unique groups into account to avoid situations like https://github.com/coiled/benchmarks/issues/1376.

Start by reading dask_expr/_groupby.py at the linked lines to understand the current split_out calculation. Review the linked benchmark issue for the motivating situation and investigate how unique-group counts could be estimated. Done means the default accounts for that estimate while preserving correct groupby aggregation behavior.

Written by the indexing model from the issue text.