Three CUDA-export memory optimizations:

- tq4_sdpa: add BLOCK_N=16 (and a BLOCK_M=32) autotune config. The superset
  is kept for big-shared-memory GPUs (A100/H100); the Triton autotuner
  auto-prunes configs that exceed a GPU's shared memory (OutOfResources ->
  inf), so the same config list also works on the 5090 (Blackwell, ~101 KB
  SMEM) where the previous smallest config did not fit.

- int4_dispatch: chunk the inline _dequant_matmul along N for vocab-sized
  weights (N>65536, i.e. only the lm_head). Avoids transiently materializing
  the full ~10 GiB bf16 lm_head when AOTI executes the int4_plain_mm custom
  op during autotune / cpp_wrapper. The runtime decode path uses the C++ dp4a
  shim and the M>4 prefill inline path is below the threshold, so this never
  enters the runtime graph -> zero runtime / accuracy impact. Applied
  unconditionally (no flag).

- cuda_backend / aoti_backend: skip occupying the GPU with the KV-cache
  buffers during AOTI compile (gated behind low_memory_mode). A new
  move_program_to_device hook places KV constants on the target device but
  immediately frees their storage (resize_(0)), so the fake-tensor device
  check passes while no real KV bytes sit on the GPU during autotune. The
  emptied buffers are re-synthesized as zeros at the _unlift_graph clone and
  at serialization, and excluded from constant dedup (resize_(0) gives every
  KV data_ptr 0, which would otherwise collapse same-shape caches across
  layers).

Result on 2xA100: Gemma4-31B @128k no-TQ export peak 36.3 -> 27.0 GiB; the
exported model runs correctly (output "...Paris.").

…nly frees genuinely all-zero kv_cache.* buffers (count_nonzero==0); preserves TQ4 centroids/boundaries/rotation/rotation_T