[DLPack] Update to include DLPack C Exchange API

[tqchen]

This PR updates to include dlpack C exchange api for fast exchange at c-extension level without going through python.

[tqchen]

Respective dlpack PR on python spec dmlc/dlpack#180

[tqchen]

cc @rgommers @leofang

[tqchen]

Implements #973

[tqchen]

cc @rgommers @leofang i think this should be ready to merge

[rgommers]

Thanks @tqchen! I refreshed my memory here, I think this PR largely looks good and we're probably ready to put this in. My main question right now is: given it's been in use in PyTorch for a couple of months now, do you have any real-world experience to share so far? No issues, did it help a lot performance-wise, etc.

[tqchen]

thanks @rgommers , updated to address your reviews. In terms of the current state, it has been merged into PyTorch and used through some of the tvm-ffi integrations through paddle and other env. So far no issues identified and it does help reducing host overhead in the specific case of tvm-ffi

[rgommers]

We touched on this PR in the community meeting yesterday. The general feeling was that it's time to merge this. @seberg and @leofang may want to have another look. I plan to merge this at the end of next week unless there are more concerns.

The feeling in the meeting is that it would be better if this new C API in DLPack would be a complete replacement for the __dlpack__ Python API. Right now it's a partial replacement; e.g., one thing that is missing is the "copy to CPU" functionality. That said, the C struct can be extended in future DLPack releases to add more functionality. And this can be done in a backwards-compatible fashion, and has no impact on what the array API standard says about __dlpack_c_exchange_api__. So while not completely optimal, no one objected to merging this PR in its current state.

[rgommers]

Thanks @tqchen. LGTM now, and CI is happy. I've put a reminder in my calendar to merge this by the end of next week unless there are new unresolved comments.