Use CUDA virtual memory for pinned memory allocation. (#10850)

- Add a grow-only virtual memory allocator.
- Define a driver API wrapper. Split up the runtime API wrapper.

tests/cpp/metric/test_distributed_metric.cc

@@ -47,7 +47,7 @@ class TestDistributedMetric : public ::testing::TestWithParam<Param> {
 
     std::int32_t n_workers{0};
     if (device.IsCUDA()) {
-      n_workers = common::AllVisibleGPUs();
+      n_workers = curt::AllVisibleGPUs();
     } else {
       n_workers = std::min(static_cast<std::int32_t>(std::thread::hardware_concurrency()), 3);
     }