Make it easier to run aggregations over nested elements in nf.eval, nf.query and nf.nested.nest

[hombit]

Feature request

Today, we have these ways to aggregate a single nested column values:

• nf.reduce(np.mean, "lc.mag") - good, but not cheap and requires to join the output back to the frame
• nf.eval("lc.mag.groupby(by=lc.mag.index).mean()") - expansive and not intuitive

It would be nice if we can develop an easier way of doing such aggregations. Options I see:

1. Currently, we can do nf.eval("lc.mag.mean()") / nf["lc.mag"].mean(), but it would output the aggregation over all the flat values, which is, especially in the first case, not intuitive. We can redefine it.
2. Add special interface for nested aggregations with .nest accessor, e.g. nf.lc.nest.mean() would return nf.shape[0] mean values.
3. Add special methods which would work in eval/query environment only, e.g. nf.eval("lc.mag.nest_mean()")

However I'm not sure how we'd make all these performant, it looks like pyarrow provides almost zero tooling for that. Maybe we can use things like numpy.ufunc.reduceat and scipy.ndimage.mean.

Before submitting
Please check the following:

• I have described the purpose of the suggested change, specifying what I need the enhancement to accomplish, i.e. what problem it solves.
• I have included any relevant links, screenshots, environment information, and data relevant to implementing the requested feature, as well as pseudocode for how I want to access the new functionality.
• If I have ideas for how the new feature could be implemented, I have provided explanations and/or pseudocode and/or task lists for the steps.

[hombit]

Some motivating benchmarks

from nested_pandas.datasets import generate_data

nf = generate_data(10_000, 1000)

%timeit nf.reduce(np.mean, 'nested.flux')
# 43.3 ms ± 224 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)

flux = pa.array(nf['nested']).field('flux')  # this is fast, ~ 5μs
%timeit np.add.reduceat(flux.values, flux.offsets[:-1]) / np.diff(flux.offsets)
1.92 ms ± 34.6 µs per loop (mean ± std. dev. of 7 runs, 1,000 loops each)