[MLIR] Guard Cuda 12.0+ newer driver APIs with CUDA_VERSION macro checks

[guraypp]

Summary

Fixes #64529
https://github.com/llvm/llvm-project/issues/64529

[guraypp]

This is a low tech solution to fix the problem fast in github.

If this runtime is going to be used, we perhaps need something more complex, we might add stub functions, or we check driver version using API. If we get to that point, we'll think of something more appropriate.

[mehdi_amini]

Nit:

• Commit title may be more descriptive as "Guard Cuda 12.0+ newer driver APIs behind CUDA_VERSION macro checks"
• your commit message should probably say Fixes #64529 instead of "Address" so that it auto-close the issue.
• We're using GitHub Pull-requests now :)

mlir/lib/ExecutionEngine/CudaRuntimeWrappers.cpp

mgpuMemHostUnregisterMemRef(int64_t rank,
   auto *ptr = descriptor->data + descriptor->offset * elementSizeBytes;
   mgpuMemHostUnregister(ptr);
 }
 
 extern "C" MLIR_CUDA_WRAPPERS_EXPORT void mgpuSetDefaultDevice(int32_t device) {
   defaultDevice = device;
 }
 
+///
+/// Runtime methods using CUDA 12.0+ driver
+///
+
+#if (CUDA_VERSION >= 12000)
+
 extern "C" MLIR_CUDA_WRAPPERS_EXPORT void mgpuTensorMapEncodeTiled(
     CUtensorMap *tensorMap,             // Tensor map object
     CUtensorMapDataType tensorDataType, // Tensor data type
     cuuint32_t tensorRank,              // Dimensionality of tensor
     void *globalAddress,                // Starting address
     const cuuint64_t *globalDim,        // Tensor size (number of elements)
     const cuuint64_t *globalStrides,    // Stride size (in bytes)
     const cuuint32_t *boxDim,           // Traversal box (number of elments)
) {
   // Copy created tensor map to device
   CUdeviceptr dTensorMap;
   CUDA_REPORT_IF_ERROR(cuMemAlloc(&dTensorMap, sizeof(CUtensorMap)));
   CUDA_REPORT_IF_ERROR(cuMemcpy(dTensorMap,
                                 reinterpret_cast<CUdeviceptr>(&tensorMap),
                                 sizeof(CUtensorMap)));
   return reinterpret_cast<void *>(dTensorMap);
 }
+#endif
 
 #ifdef MLIR_ENABLE_CUDA_CUSPARSE
 
 ///
 /// Wrapper methods for the cuSparse library.
 ///
 
 // Some macro magic to get float/double alpha and beta on host.