feat(cudf): Disable async MR priming in Velox cuDF (#14997)

Summary:
This PR skips the default "priming" step of the RMM async memory resource. This reduces initialization costs and has benefits for multi-process applications.

xref:
- rapidsai/rmm#2060
- rapidsai/rmm#1931
- rapidsai/rmm#2051

Pull Request resolved: #14997

Reviewed By: Yuhta

Differential Revision: D83668107

Pulled By: kgpai

fbshipit-source-id: 41b6bd5807f60b0e1c76a1c91dd26f7e5451255a

Author: bdice

velox/experimental/cudf/exec/Utilities.cpp

@@ -50,9 +50,8 @@ namespace {
       makeCudaMr(), rmm::percent_of_free_device_memory(percent));
 }
 
-[[nodiscard]] auto makeAsyncMr(int percent) {
-  return std::make_shared<rmm::mr::cuda_async_memory_resource>(
-      rmm::percent_of_free_device_memory(percent));
+[[nodiscard]] auto makeAsyncMr() {
+  return std::make_shared<rmm::mr::cuda_async_memory_resource>();
 }
 
 [[nodiscard]] auto makeManagedMr() {
@@ -78,7 +77,7 @@ std::shared_ptr<rmm::mr::device_memory_resource> createMemoryResource(
   if (mode == "pool")
     return makePoolMr(percent);
   if (mode == "async")
-    return makeAsyncMr(percent);
+    return makeAsyncMr();
   if (mode == "arena")
     return makeArenaMr(percent);
   if (mode == "managed")