fix: return 0-D array for full reductions per Array API standard

[Abineshabee]

Fixes #921

Problem

sparse.sum() and all other full reductions (max, min, prod, mean, any, all, einsum) were returning a NumPy scalar instead of a 0-D array, violating the Array API standard.

Root Cause

SparseArray.reduce() had if out.ndim == 0: return out[()] which unwrapped the 0-D COO into a scalar. _einsum_single() also had two explicit scalar return paths.

Changes

• _sparse_array.py — remove scalar unwrap in reduce(), unify mean() path
• _common.py — fix both scalar paths in _einsum_single()
• test_einsum.py — update scalar-output branches to use .todense()
• test_array_function.py — add TestArrayAPIReductions with 16 regression tests

Testing

6079 passed, 0 failed

[hameerabbasi]

One thing that can be improved here: out[()] itself should just return out, never a NumPy scalar.

[Abineshabee]

Fixed out[()] scalar unwrap — reduce() and var() now return 0-D sparse arrays directly per Array API standard. All 316 tests passing. @hameerabbasi please re-review!

[Abineshabee]

Please let me know if any further changes are needed!

[hameerabbasi]

Just one small concern.

[hameerabbasi]

Huh? s.mean() should produce SOME output.

[codspeed-hq]

Merging this PR will degrade performance by 54.75%

⚠️ Different runtime environments detected

Some benchmarks with significant performance changes were compared across different runtime environments,
which may affect the accuracy of the results.

Open the report in CodSpeed to investigate

❌ 1 regressed benchmark
✅ 339 untouched benchmarks

⚠️ Please fix the performance issues or acknowledge them on CodSpeed.

Performance Changes

	Benchmark	BASE	HEAD	Efficiency
❌	test_index_slice[side=100-rank=2-format='gcxs']	1.7 ms	3.8 ms	-54.75%

---

_{Comparing Abineshabee:fix/array-api-reduction-returns-0d-array (9304ca7) with main (3589a7c)}