REF: Reductions

[jbrockmendel]

We have reductions implemented in nanops, _libs.groupby, and _libs.window.aggregations. We should refactor these with the following goals in mind:

1. Have one/fewer distinct implementations
2. Avoid copies, particularly in the nanops versions where we do something like values[notna(values)]
3. Chunked-friendliness, so that we can re-write ArrowExtensionArray._groupby_op to operate chunk-by-chunk, avoiding a copy in multi-chunk cases. (This could also be useful for hypothetical distributed EAs)
4. Avoid casting/inference in nanops
5. update Do axis=1 reductions without transposing/copying, inspired by PERF: axis=1 reductions with EA dtypes #54341

The implementation of group_skew is derived from https://www.johndcook.com/blog/skewness_kurtosis/ which includes a method for "adding" multiple RunningStats instances. Something like that could be adapted for 3).

[mroeschke]

Just noting that _libs.window.aggregations should ideally keep its implementation since the sliding window aggregation is performant sensitive. I think the other reductions could be implemented in terms of the sliding windowing aggregation i.e. they would be non-overlapping windows