CLN/PERF: no need for kahan for int group_cumsum

[mzeitlin11]

Surprised at lack of impact here - doesn't noticeably affect benchmarks.

Targeting the cython algo specifically shows an improvement (but smaller than I'd expect given the removed operations):

import numpy as np
import pandas._libs.groupby as libgroupby

N = 4_000_000
vals = np.random.randint(0, 10, (N, 5), dtype=np.int64)
result = np.empty_like(vals)

%timeit libgroupby.group_cumsum(result, vals, np.ones(N, dtype="int"), 1, False)
# 28.9 ms ± 805 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)  # this pr
# 37.3 ms ± 206 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)   # master

[mzeitlin11]

Doubt this affects compiled result, but may as well to not depend on a smart compiler avoiding this extra indexing step

[jreback]

umm this is affecting all dtypes. do we not have tests for this for small floats?

[mzeitlin11]

This is inside the else block from an if statement if numeric == float32_t or numeric == float64_t: so only non-floats should end up here.

[jreback]

ahh ok that was not clear from the difff

[jreback]

can you add a comment to that effect (maybe just on the float32/64 branch, e.g. using Kahan summation)

[mzeitlin11]

Have added a comment

[jreback]

hmm seemingly unrelated failures. maybe on master? cc @jbrockmendel

[jreback]

thanks @mzeitlin11