Ensure nvbench initializes nvml context when built statically

cpp/cmake/thirdparty/patches/nvbench_nvml_setup.diff

@@ -0,0 +1,39 @@
+diff --git a/nvbench/device_info.cu b/nvbench/device_info.cu
+index 8edad22..a09def3 100644
+--- a/nvbench/device_info.cu
++++ b/nvbench/device_info.cu
+@@ -25,6 +25,10 @@
+
+ #include <cuda_runtime_api.h>
+
++#ifdef NVBENCH_HAS_NVML
++#include <nvbench/internal/nvml.cxx>
++#endif
++
+ #define UNUSED(x) (void)(x)
+
+ namespace nvbench
+@@ -47,6 +51,10 @@ device_info::device_info(int id)
+   NVBENCH_CUDA_CALL(cudaGetDeviceProperties(&m_prop, m_id));
+
+ #ifdef NVBENCH_HAS_NVML
++  // NVML's lifetime should extend for the entirety of the process, so store in a
++  // global.
++  [[maybe_unused]] static auto nvml_lifetime = NVMLLifetimeManager();
++
+   // Retrieve the current device's pci_id as a null-terminated string.
+   // Docs say 13 chars should always be sufficient.
+   constexpr int pci_id_len = 13;
+diff --git a/nvbench/internal/nvml.cxx b/nvbench/internal/nvml.cxx
+index 4f750bc..ce7d874 100644
+--- a/nvbench/internal/nvml.cxx
++++ b/nvbench/internal/nvml.cxx
+@@ -64,8 +64,4 @@ private:
+   bool m_inited{false};
+ };
+
+-// NVML's lifetime should extend for the entirety of the process, so store in a
+-// global.
+-auto nvml_lifetime = NVMLLifetimeManager{};
+-
+ } // namespace

cpp/cmake/thirdparty/patches/nvbench_override.json

@@ -7,6 +7,11 @@
           "file" : "${current_json_dir}/nvbench_global_setup.diff",
           "issue" : "Fix add support for global setup to initialize RMM in nvbench [https://github.com/NVIDIA/nvbench/pull/123]",
           "fixed_in" : ""
+        },
+        {
+          "file" : "${current_json_dir}/nvbench_nvml_setup.diff",
+          "issue" : "Add support for nvml with static nvbench [https://github.com/NVIDIA/nvbench/pull/148]",
+          "fixed_in" : ""
         }
       ]
     }