feat: add a dtype= accumulator to ak.sum (overflow-safe, copy-free integer sums)

[ianna]

Motivation

Integer reductions in the descriptive statistics overflow because the sum accumulates at the input dtype in the segmented kernel. The Python-level fix (#4223) casts integer input to float64 up front, which allocates a full-size temporary — the copy raised against #3527. NumPy avoids this by casting inside the reduction (np.add.reduce(x, dtype=np.float64)); awkward couldn't, because it had no float64-accumulator / integer-input reduce kernel.

This PR adds that primitive: ak.sum(x, dtype=np.float64) sums integer/bool input directly into float64, matching NumPy and avoiding a promoted copy.

Changes

• Kernels: 8 new awkward_reduce_sum_float64_{int,uint}{8,16,32,64}_64 specializations in kernel-specification.yml and registered in awkward-cpp/src/cpu-kernels/awkward_reduce_sum.cpp. The kernel body is unchanged — awkward_reduce_sum already casts per element via static_cast<OUT>, so a float64 accumulator over integer input is a pure specialization. Generated kernels.h and the signature tables are regenerated from the spec.
• Sum reducer (_reducers.py): Sum(dtype=None). When set, the accumulator/output dtype flows into the segmented kernel selection and the axis=None path; the bool path casts its small result.
• nplike: dtype= added to sum on all four backends (array_module, cupy, typetracer, and the protocol in numpy_like), forwarding to NumPy/CuPy sum(dtype=).
• ak.sum: new dtype= keyword + docstring.
• Tests: tests/test_4223_sum_dtype_accumulator.py.

Scope / follow-ups

• ak.mean/ak.var are intentionally not rewired here. Doing so would hard-couple the pure-Python package to the new kernel — integer mean/var over an axis would break against any released awkward-cpp predating it. They should adopt dtype= (and drop the fix: overflow-safe stats without needless copies #4223 promotion) in a follow-up, once these kernels ship in a released awkward-cpp.
• This is Phase 1 of the plan for copy-free overflow-safe stats; var/std additionally need a fused sum-of-squares reducer (Phase 2).

Release ordering

Kernels live in awkward-cpp; the Python wiring is in awkward. The axis=None path works against any awkward-cpp (it maps onto NumPy/CuPy sum(dtype=)), but the segmented axis != None path requires a rebuilt awkward-cpp — so awkward-cpp must be released with these kernels before awkward relies on them. The two segmented tests skip when the kernels are absent.

Non-goals

ak.sum/ak.prod default dtype semantics are unchanged (NumPy preserves integer dtype); the float64 accumulator is opt-in. Complex/datetime unaffected.

AI disclosure

Parts of this PR were developed with AI assistance (per CONTRIBUTING.md).

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 84.05%. Comparing base (802b733) to head (cc9b934).
✅ All tests successful. No failed tests found.

Additional details and impacted files

Files with missing lines	Coverage Δ
src/awkward/_nplikes/array_module.py	93.71% <100.00%> (ø)
src/awkward/_nplikes/cupy.py	91.45% <100.00%> (ø)
src/awkward/_nplikes/numpy_like.py	100.00% <ø> (ø)
src/awkward/_nplikes/typetracer.py	77.38% <100.00%> (+0.14%)	⬆️
src/awkward/_reducers.py	99.03% <100.00%> (+0.71%)	⬆️
src/awkward/operations/ak_sum.py	94.59% <100.00%> (ø)

[github-actions]

The documentation preview is ready to be viewed at http://preview.awkward-array.org.s3-website.us-east-1.amazonaws.com/PR4232