The f16 gemv_bt_via_f32 path extended cfg gates to include aarch64 but
simd_dot_f32 only had an x86_64 implementation, causing a compile error
on Apple Silicon and other aarch64 targets. Add a NEON implementation
using dual vfmaq_f32 accumulators to hide FMA latency, matching the
pattern used by the AVX2 path.