nx_q4k_matmul_bench.nx

nx_q4k_matmul_bench.nx -- fused vs materialize-then-dot bench. Compares the two paths for a K=2048 row × col dot product (8 Q4_K super-blocks = ~3 KB of quantized weights): Path A: nx_gguf_dequant_q4_k(buf, ..., scratch_2048) + manual scalar dot product over scratch_2048 vs col_2048 Path B: nx_q4k_dot_row_col(buf, ..., col_2048, iter) fused Honest framing: numbers are on qemu-riscv64 emulation; this is OUR own two paths compared, NOT a claim against ggml/llama.cpp/CUDA. Comparison against external frameworks requires running them on the same hardware.