Discussion: Implement "lower-level" APIs for cuda.parallel that do not accept array inputs?

[leofang]

we could consider changing the API to not accept __cuda_array_interface__ objects, and instead have the user pass in the required information (pointer, size, dtype, etc.,). This allows each library/user to compute that information in the most efficient way possible rather than making it our responsibility.

Let's have a separate issue to track this. Thinking about this more we should try to make the current (low-level) interface look more like a 1:1 binding to the bare C++ one. This is what we do for cuda.cooperative too. Pythonic interface can come later.

Originally posted by @leofang in #3718 (comment)