[BUG] All_reduce over >INT_MAX elements hits RuntimeError

**Describe the bug** CUDA global reductions fail when the input has more than `INT_MAX` elements. **To Reproduce** ```python import mlx.core as mx a = mx.ones((32768, 32768, 2), dtype=mx.float16) mx.eval(mx.sum(a)) ``` This fails with: `RuntimeError: cudaGraphAddKernelNode(&node, graph_, NULL, 0, &params) failed: invalid argument` **Expected behavior** This reduction should complete and return a scalar. **Desktop (please complete the following information):** OS Version: Ubuntu 22.04.5 LTS Version: mlx 0.31.1, mlx-cuda-13 0.31.1 GPU: NVIDIA RTX 6000 Ada Generation, driver 580.126.09 **Additional context** 32768 * 32768 * 2 == 2**31, which is one element past INT_MAX. Other reductions such as prod, min, max, all, and any hit the same issue.