[Python] group_by functionality directly on large dataset, instead of on a table?

I have a large dataset that I would like to use `group_by` on without having to read the entire table into memory first.

After reading the documentation it seems `dataset.to_batches` is the best way of doing this? But it gets really complex when using other aggregation methods than for example `count` and `sum`.

I implemented it like below for `count` and `sum`, but for other more complex aggregations I am still forced to read the entire table.

```python
table = []
for batch in ds.to_batches(columns=columns, filter=filters, batch_size=1e6):
t = pyarrow.Table.from_batches([batch])
table.append(t.group_by(group_by).aggregate(agg))
table = pyarrow.concat_tables(table)

# then after this I use group_by again on the concatenated table with `sum` as aggregation method
```

Thankful for any pointers or comments!

Start by reviewing the Python Dataset.to_batches and Table.group_by APIs, focusing on how aggregations are handled batch by batch. Define what a direct large-dataset group_by should support beyond count and sum, and consider how completion can avoid materializing the entire table in memory.

Written by the indexing model from the issue text.