nx_simd_vdot_test.nx

nx_simd_vdot_test.nx -- widening dot product i16x16 -> i64. The kernel of every ML inference inner loop, every convolution stride, every audio/DSP filter on x86. vpmaddwd does (a*b) pair-wise into 8 i32 lanes in ONE instruction; RV-V does it via vwmul.vv + vwredsum.vs (widening at the codegen layer). We test the bit-exact behaviour vs a hand-written scalar reduce because the widening prevents intermediate overflow that the naive vmul+vreduce_sum would suffer with values > 181 (since 181*181 = 32761 ~ INT16_MAX).