Make DLPack a submodule?

[leofang]

I am looking into DLPack as we likely need to fix it for ROCm (it supports the kDLROCM option). One thing I noticed is that we basically reimplemented everything from the DLPack header dlpack.h in cupy/core/dlpack.pyx. I suppose like Jitify it's also a single header thing, so we might as well make it a submodule?

This will also prepare us for the Python Array API involvement, for which DLPack is being considered as the data exchange format (data-apis/consortium-feedback#1; see also the Array API Standard draft), but it currently lacks

1. support for complex numbers (Support complex numbers in DLPack protocol dmlc/dlpack#50)
2. specification for synchronization semantics as in CUDA Array Interface v3 (Specify synchronization semantics dmlc/dlpack#57)

so it will likely be revised in the near future. If this is acceptable then please assign it to me.

[leofang]

UPDATE: We bundle dlpack.h directly without making it a submodule.