[CUDA] Reduce use of managed memory (#2725)

* Use async cuda malloc managed with cuda 13

* add pool threshold

* refactor for regular cuda malloc

* load eval gpu for cuda

* remove use of cuda pool, use cuda free async

* fix

* fix

* fix

* fix

* fix + comment

mlx/backend/cuda/arange.cu

@@ -41,9 +41,8 @@ void Arange::eval_gpu(const std::vector<array>& inputs, array& out) {
   if (out.size() == 0) {
     return;
   }
-  out.set_data(allocator::malloc(out.nbytes()));
-
   auto& encoder = cu::get_command_encoder(stream());
+  out.set_data(cu::malloc_async(out.nbytes(), encoder.stream()));
   encoder.set_output_array(out);
 
   dispatch_int_float_types(out.dtype(), "Arange", [&](auto type_tag) {
@@ -58,7 +57,7 @@ void Arange::eval_gpu(const std::vector<array>& inputs, array& out) {
           num_blocks,
           block_dims,
           0,
-          out.data<OutType>(),
+          gpu_ptr<OutType>(out),
           out.data_size(),
           static_cast<CTYPE>(start_),
           static_cast<CTYPE>(start_ + step_) - static_cast<CTYPE>(start_));