Should IPyBuffer support Memory<T> and therefore Tensor<T>?

[tonybaloney]

When looking at the new MEAI Embedding, the constructor takes ReadOnlyMemory.

When a method returns a PyBuffer which contains a vector of single, double or other vector type creating an Embedding<T> requires copying the values to an array then creating an Embedding from that array.

1. Does it make sense to expose AsMemory<T> methods for PyBuffer?
2. What considerations do we need to have on the reference count for the memory?
3. Would this be more efficient than the existing implementation?

https://learn.microsoft.com/en-us/dotnet/standard/memory-and-spans/memory-t-usage-guidelines

See example

[atifaziz]

See #504 where I've laid some concerns with the existing API and a proposal in #505 that also demonstrates how Memory<T> would work (with safety remaining to be the primary concern).