Race condition when MKLDNN is enabled

[eric-haibin-lin]

When MKLDNN is enabled, the output of any MKLDNN op will have MKL layout. If such ndarray is passed to any operator registered with FComputeEx interface, calling NDArray::data will trigger "Reorder2Default" which may cause race condition since it doesn't prevent others from writing to it at the same time.

The number of non-MKL ops registered with FComputeEx is over 50. This should be fixed and explicitly tested.

[zheng-da]

This potential race condition is very difficult to produce. It may happen if the output of an MKLDNN operator is used with a sparse array in an operator. The operator has to perform computation on these two arrays (an MKLDNN array and a sparse array) with FComputeEx. However, the current implementation of sparse operators doesn't support this case and fall back using FCompute.

[zheng-da]

@eric-haibin-lin i think you can close this issue now.

[eric-haibin-lin]

could you reference the PR?