nx_dequant_iter_bench.nx

nx_dequant_iter_bench.nx -- partial-access throughput comparison. Measures: when a caller needs only K of 256 values per Q4_K block, what's the substrate-side cost? Two paths timed: A) WHOLE-BLOCK PATH: nx_gguf_dequant_q4_k materializes all 256 values for every block, even if caller throws 256-K of them away. B) ITER PATH: nx_q4k_iter_init once per block + K calls to nx_q4k_iter_value for just the K needed values. Workload: N blocks * K reads each, for K in {1, 8, 32, 256}. HONEST FRAMING: - These numbers are on qemu-riscv64 emulation; native silicon would be faster. - The comparison is BETWEEN OUR OWN two paths, not against any other framework. No "Nx faster than CUDA / llama.cpp" claim. - At K=256 the iter path is expected to be roughly equivalent to (or slightly slower than) the whole-block path -- both do the same total work. At K<256 the iter path skips work.