feat: I32x16 Base17 ops — from_i16_slice, abs, to_i16_array, cmpge_zero_mask Added to all 3 tiers (AVX-512 / AVX2 / scalar): from_i16_slice(&[i16]) → I32x16 — load 16×i16, sign-extend to 16×i32 abs() → I32x16 — absolute value per lane to_i16_array() → [i16; 16] — narrow 16×i32 back to 16×i16 cmpge_zero_mask() → u16 — bit mask where lane >= 0 These are the primitives bgz17_bridge.rs needs to replace its 92 raw intrinsics with crate::simd::I32x16 calls. Fixed duplicate abs() in AVX-512 I32x16. 19 bgz17_bridge tests pass. https://claude.ai/code/session_01ChLvBfpJS8dQhHxRD4pYNp #77