CUDA backend: compile (#2276)

* CUDA backend: compile

* Rename kernels/ to device/

mlx/backend/cuda/device/config.h

@@ -0,0 +1,12 @@
+// Copyright © 2025 Apple Inc.
+
+// This file is used by both CUDA kernel code and host-only C++ code.
+
+#pragma once
+
+// The maximum dimensions of shape/strides passed as kernel parameters.
+#define MAX_NDIM 8
+
+// All existing NVIDIA hardware has a fixed 32 warp size. Though a built-in
+// warpSize variable exists, using it would prevent compile-time optimizations.
+#define WARP_SIZE 32