Design the TensorMap object

[leofang]

I am not sure if we even need a class full of states, attributes, and methods to wrap CUtensorMap. I suspect we just need a lightweight class object + a free, factory function that takes a variety of configs and returns the object (i.e. no class constructor).

Unanswered question: Is CUtensorMap bound to a CUDA context?

[leofang]

Is CUtensorMap bound to a CUDA context?

Asked internally and the answer is yes. It means the TensorMap object would need to offer a reverse lookup to device/context, as done in other objects.