Allow any backing array type and add Adapt support

[darsnack]

Closes #8. Currently, I didn't add tests since the repo doesn't have a build kite pipeline. Does JuliaReinforcementLearning have the app installed? I'm happy to help coordinate this with JuliaGPU if needed.

[darsnack]

With some testing on a machine with a GPU, I see that this isn't going to just work. Let me play around with it some more.

[findmyway]

I think push!, pushfirst!, pop! or append! should just work. The problem is with views and element-wise operations. Could you share your desired usage? I guess a convert function to convert a view into a CuArray would meet most cases?

[darsnack]

I had to make some minor adjustments to the code to get push! etc. to work. I also added tests that replicate the full test suite for CuArrays. This will not be able to run on GHA, so tests will fail right now. I can verify that they pass locally on my machine with a GPU.

For element-wise operations, broadcasting is the recommended solution for GPUs. In this case, we will need to take on a CUDA.jl dep to specify the broadcasting style. The code is a one-liner, but I'll leave it to you whether the dep is worth it. I don't think this would be a common use case, and a user can always index the full array.

For views, it becomes an issue because we have a SubArray of a CircularArrayBuffer which CUDA.jl will treat as non-isbits. We could return a view of the underlying buffer instead which would only be an issue if the indices are non-contiguous (triggering scalar indexing). This is an outstanding issue for plain CuArrays as well.

My use-case is that I have data streaming in as CuArrays and I need to store these in a multi-dimensional circular buffer for use later. I am taking a GPU-compatible view of the multi-dimensional circular buffer later in the code and doing some computation with it. Not having a CuArray as the underlying buffer means copying data back and forth from the GPU, and my reason for using this package over DataStuctures is that it minimizes the allocations/copies in my use case (both for CPU and GPU).

[findmyway]

For element-wise operations, broadcasting is the recommended solution for GPUs. In this case, we will need to take on a CUDA.jl dep to specify the broadcasting style. The code is a one-liner, but I'll leave it to you whether the dep is worth it. I don't think this would be a common use case, and a user can always index the full array.

👍