[SYCL][CUDA] Improve error message for exceeding CUDA grid limits (#4563)

Improves the error message when the user specifies a number of work groups in certain dimension, which exceeds CUDA's max grid dimension limits.

Before this PR the error message is not very informative:

```
PI CUDA ERROR:
        Value:           1
        Name:            CUDA_ERROR_INVALID_VALUE
        Description:     invalid argument
        Function:        cuda_piEnqueueKernelLaunch
        Source Location: /home/tadej/llvm/sycl/plugins/cuda/pi_cuda.cpp:2662

terminate called after throwing an instance of 'cl::sycl::runtime_error'
  what():  Native API failed. Native API returns: -30 (CL_INVALID_VALUE) -30 (CL_INVALID_VALUE)
This is especially confusing when porting code from CUDA, due to the fact that dimensions in SYCL are flipped compared to CUDA and CUDA has different limits for different dimensions.
```

Tested by: intel/llvm-test-suite#952

Author: t4c1

sycl/source/detail/error_handling/enqueue_kernel.cpp

@@ -268,6 +268,31 @@ bool handleInvalidWorkItemSize(const device_impl &DeviceImpl,
   return 0;
 }
 
+bool handleInvalidValue(const device_impl &DeviceImpl,
+                        const NDRDescT &NDRDesc) {
+  const plugin &Plugin = DeviceImpl.getPlugin();
+  RT::PiDevice Device = DeviceImpl.getHandleRef();
+
+  size_t MaxNWGs[] = {0, 0, 0};
+  Plugin.call<PiApiKind::piDeviceGetInfo>(
+      Device, PI_EXT_ONEAPI_DEVICE_INFO_MAX_WORK_GROUPS_3D, sizeof(MaxNWGs),
+      &MaxNWGs, nullptr);
+  for (unsigned int I = 0; I < NDRDesc.Dims; I++) {
+    size_t NWgs = NDRDesc.GlobalSize[I] / NDRDesc.LocalSize[I];
+    if (NWgs > MaxNWGs[I])
+      throw sycl::nd_range_error(
+          "Number of work-groups exceed limit for dimension " +
+              std::to_string(I) + " : " + std::to_string(NWgs) + " > " +
+              std::to_string(MaxNWGs[I]),
+          PI_INVALID_VALUE);
+  }
+
+  // fallback
+  constexpr pi_result Error = PI_INVALID_VALUE;
+  throw runtime_error(
+      "Native API failed. Native API returns: " + codeToString(Error), Error);
+}
+
 bool handleError(pi_result Error, const device_impl &DeviceImpl,
                  pi_kernel Kernel, const NDRDescT &NDRDesc) {
   assert(Error != PI_SUCCESS &&
@@ -315,6 +340,9 @@ bool handleError(pi_result Error, const device_impl &DeviceImpl,
         "slice pitch) are not supported by device associated with queue",
         PI_INVALID_IMAGE_SIZE);
 
+  case PI_INVALID_VALUE:
+    return handleInvalidValue(DeviceImpl, NDRDesc);
+
     // TODO: Handle other error codes
 
   default: