[NDArray] Set NDArray::Container.shape_ in NDArray::FromDLPack

[hlu1]

In some cases, the shape info in DLTensor.shape might not be dynamically allocated and may cease to exist after passing the DLTensor to NDArray. We can avoid this problem by setting NDArray::Container.shape_ at construction time and assign it back to DLTensor.shape.

[tqchen]

THanks @hlu1 however, according to the DLPack conventionm, the DLManagedTensor won't be deleted until the deleter is called, could it due to the exporter that violates the convention?

[hlu1]

I'm referring to the case that DLManagedTensor.dl_tensor.shape which is an int64_t* could be pointing to an array that's allocated in the stack and not on the heap. People are usually pretty careful to make sure the data is allocated on the heap and properly deleted in the deleter, but forget to do the same for the shape info.

[hlu1]

Another thing that's related:

std::vector<int64_t> NDArray::Shape() const {
  return get_mutable()->shape_;
}

https://github.com/apache/incubator-tvm/blob/master/src/runtime/ndarray.cc#L251-L253

If we don't set shape_ in NDArray::FromDLPack, we'll need to fix NDArray::Shape() to make sure it returns the right shape info.

[tqchen]

get you, Thanks @hlu1 !