Can we contextualize `parent_type`?

首先阅读链接的 PR，以及 issue 中关于 parent(x)、parent_type(x) 和所提议的 BufferStyle 类别的讨论。比较所述示例——SubArray、Tridiagonal 和 FillArrays——并确定是否可以指定一个具体的 trait 或通用的 f(x) = f(parent(x)) 设计。完成的标准是形成一个确定且可实现的接口提案。

When I made this PR to base some good points were brought up about what we knew about parent(x) compared to x. With a few exceptions (that we should probably clean up), we simply use parent_type as convenient syntax for communicating the type of the underlying data buffer/collection. An informal search through juliahub shows that a handful of packages define something very similar independent of this package, so it's clearly something that has utility already. I'd like to know if we could do better.

The biggest hangup here is how we succinctly describe what assumptions persists between x and parent(x). The only approachable idea I have is a method like is_lazy_wrapper when we can assume that the wrapper type changes nothing. However, that seems extremely limiting and questionably useful.

A more approachable aspect to this is describing the kind of type that is parent_type(x). For example, sub-typing something like BufferStyle with these subtypes:

• WrappedBuffer: single wrapped buffer like SubArray.
• MultiBuffer: multiple distinct buffers like Tridiagonal
• ComputedBuffer: buffer is computed on the fly and isn't stored in memory (e.g., FillArrays).

However, creating traits for describing parent_type(x) isn't terribly useful until we have a generic approach for doing f(x) = f(parent(x)).