Avoid apply in _compute_sum_of_squares for groupby std agg

[rjzamora]

This PR addresses #6034 by avoiding the apply call in _compute_sum_of_squares for std groupby aggregations.

Motivation: The following does not currently work with dask_cudf, because the dask.dataframe implementation uses groupby-apply (which does not exhibit "correct" behavior for all cases in cudf):

import cudf, dask_cudf

df = cudf.DataFrame({'a': [1,1,2,2],'b': [4,5,6,10]})
ddf = dask_cudf.from_cudf(df, npartitions=2)
ddf.groupby('a').agg({'b':['mean','std']}).compute()

Note that this change also improves pandas-backed dask.dataframe performance:

import cudf, dask_cudf
import dask.dataframe as dd
import pandas as pd
import numpy as np

size = 100_000_000
df = pd.DataFrame({'a': np.random.randint(10, size=size),'b': np.random.randint(10, size=size)})
ddf = dd.from_pandas(df, npartitions=2)

%timeit ddf.groupby('a').agg({'b':['mean','std']}).compute()

Master: 3.46 s ± 71.5 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
This PR: 1.74 s ± 36.9 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

• Tests added / passed
• Passes black dask / flake8 dask

[mrocklin]

Oh great! I'm glad to see that it was possible to find a solution to this that used only more efficient Pandas/cudf API. Seeing the speedup on pandas is great, and knowing that it allows cudf operations to work at all is awesome. This seems to pass tests, so I'm +1, but I figure we wait a day in case someone like @TomAugspurger has thoughts.

[TomAugspurger]

Just one question. Looks good otherwise.

[rjzamora]

@shwina - Let me know if you have any thoughts/advice on how I am using cudf/groupby here. [EDIT: More specifically, is there a "public" attribute I should use instead?]

[rjzamora]

Oh - It looks like Ashwin pointed out the keys attribute here.

[kkraus14]

Just a clarification, we can use groupby.apply but we should avoid it for both Pandas and cuDF whenever possible as it goes through a slower iteration based code path.

[jcrist]

LGTM, thanks @rjzamora. Merging.