fix(hpc/activations): sigmoid_f32 stride mismatch (orphan rescue from PR #154)

[AdaWorldAPI]

Summary

Rescues an orphaned fix from PR #154: the Codex review of #154 flagged a stride-mismatch bug in sigmoid_f32 that I committed (b5a63fc7) but landed AFTER the merge cycle had completed — so the merged version of #154 still carries the bug.

The bug

Same-shaped contiguous views with different memory orders (C-order input + F-order output) both succeeded at as_slice_memory_order but with mismatched logical indexing. The flat SIMD primitive then wrote sigmoid values into the wrong output coordinates instead of falling back to the stride-aware Zip cold path.

Reproducer (also the regression test added in this PR):

let x: Array2<f32> = arr2(&[[0.0_f32, 100.0], [-100.0, 0.0]]); // C-order, strides [2, 1]
let mut out: Array2<f32> = Array::zeros((2, 2).f());             // F-order, strides [1, 2]
sigmoid_f32(x.view(), out.view_mut());
// Before fix: wrong values at out[[0,1]] and out[[1,0]] (logical indexing mismatch)
// After fix:  out[[0,0]]≈0.5, out[[0,1]]≈1.0, out[[1,0]]≈0.0, out[[1,1]]≈0.5

The fix

Add the same x.strides() == out.strides() guard that hpc/vml.rs already uses in dispatch_unary_contig / dispatch_binary_contig (vml's unary and binary fns were never affected because they routed through these helpers). Mismatched-stride inputs now correctly route to the stride-aware Zip cold path.

Why other W2 surfaces are unaffected

• hpc/reductions.rs (sum / mean / max / min / nrm2 / argmax / argmin) — read-only commutative/associative reductions. Memory-order of the iteration doesn't change the scalar result.
• hpc/vml.rs (16 unary + 4 binary fns) — already routed through dispatch_unary_contig / dispatch_binary_contig which carry the strides-equality guard.
• activations::softmax_f32 / log_softmax_f32 — ArrayView1 only; 1-D as_slice() returns None for non-unit stride, so the cold path engages for any strided 1-D view.

Only sigmoid_f32 was generic-D AND used the raw as_slice_memory_order calls without the guard.

Test plan

• cargo check -p ndarray --no-default-features --features std clean
• cargo test -p ndarray --lib --no-default-features --features std hpc::activations — 17 passed / 0 failed (was 16 before this fix)
• New test_sigmoid_f32_c_in_f_out_mismatched_strides exercises the exact Codex repro
• cargo fmt --all -- --check clean
• cargo clippy --no-default-features --features std -- -D warnings clean
• CI matrix

---

Generated by Claude Code