nx_gguf_load_lazy_test.nx

nx_gguf_load_lazy_test.nx -- smoke for nx_gguf_load_lazy.nx + QUANTITATIVE proof of lazy-vs-eager RAM footprint. Builds a GGUF with 2 tensors: "score" F32 [2,2] (4 values, 16 bytes raw) "weight" Q8_0 [32] (32 values, 34 bytes raw block) Test classes: A) Lazy load returns NxPlacedTensor with placement=MMAP and storage_bytes=0. B) Materialize on demand via nx_placed_tensor_get reproduces the same values nx_gguf_load_tensor would have produced. C) Demote returns to zero-storage state; re-materialize works. D) Footprint accounting: handle_bytes scales linearly with tensor count; eager would scale with tensor-DATA bytes. Verify lazy_handle_bytes(n) = n * 88. E) Not-found verdict propagated.