refactor(hpc/reductions): convert 9 pub fns to ArrayView-first (W2-1)

In-place rename per .claude/knowledge/w2-arrayview-migration.md. Each
fn now takes ArrayView<T,D> (generic-D where semantically valid;
ArrayView1<f32> for argmax/argmin which are inherently 1-D). Hot path
calls the existing SIMD primitive via as_slice_memory_order; cold path
falls back to stride-aware iter().

Tests updated: every fn gets contiguous + strided + (for generic-D)
2-D + (for Option-returning) empty coverage.

Author: claude