perf: optimize topk with heap selection, k=1 linear scan, and contiguous fast path - #36

This PR replaces the O(n log n) full-sort topk algorithm with O(n log k) binary-heap selection for moderate k, an O(n) linear scan for k=1 (greedy sampling), and a zero-copy contiguous fast path that bypasses extract_array allocation. Flank refactoring extracts a shared is_c_contiguous helper and a define_extract_view macro to eliminate code duplication.

Motivation and Context

topk() was a bottleneck in ML inference pipelines — taking up to 15ms per call (37% of a 40ms model step) while other operations ran in 1-5ms. The algorithm sorted every lane in full (O(n log n)) even when only k=1 or k=50 elements were needed. Additionally, every call allocated and copied the entire array via extract_array regardless of whether the data was already contiguous.

For a flat 152k-element array with k=50 (a real downstream sampling scenario), the full sort did ~2.6M comparisons. The heap-based approach does ~152k comparisons — a 17x algorithmic improvement, translating to roughly 18x wall-clock speedup (11.3ms → 0.63ms).

What's Changed

• Binary-heap topk selection (O(n log k)) replaces full sort when k × 4 < n
• k=1 linear scan via lane iterator — O(n) with no per-element bounds checks, eliminating sort entirely for greedy decoding
• Contiguous fast path: creates ArrayViewD directly from raw pointer instead of calling extract_array, eliminating one full-array allocation and copy per call
• is_c_contiguous helper extracted to shared location, replacing 3 duplicate implementations
• define_extract_view macro added — zero-copy immutable view extraction, symmetric with existing define_extract_view_mut
• topk.rs simplified from 796 to 373 lines via topk_axis_arm/topk_flat_arm macros

Breaking Changes

None