nx_simd_madd_probe.nx

nx_simd_madd_probe.nx -- test the x86 integer-SIMD primitive __i16x16_madd (vpmaddwd) on nx_cc_known_good. __simd_vdot_i16_x16 is RISC-V-only codegen (nx_parse.nx:1436 says nx_riscv.nx). But __i16x16_madd (nx_parse.nx:1497) is documented as x86 vpmaddwd: *acc(i32x8) += a(i16x16) . b(i16x16) pairwise. If THIS works on x86, the integer GEMM can be SIMD-accelerated here -> the whole perf-exceed unblocks. Golden: a=b=[1..16], vpmaddwd pairs -> 8 i32, sum = sum(i^2, 1..16) = 1496. license_tier: ORIGINAL