Revisit: default depth of `flatten`

[dcporter]

Following on from #9, which changed the default depth of flatten from Infinity to 1. I would like to revisit this.

A (highly unscientific) survey of friends has suggested that:

• Single-layer flattening is surprising; full flattening is the expectation.
• Use cases for not flattening all the way were not immediately obvious (although the use cases discussed were in general contrived anyway).

If the shallow flattening is undesired behavior, then the performance benefit of defaulting to depth=1 is illusory. If there's only one layer of arrays, then it's the same performance, and if there are more, then it's the wrong outcome.

Are we wrong about real-life use cases strongly favoring full flattening? If not, or even if only slightly, then I find the element of surprise to outweigh the default performance profile.

[bakkot]

As long as we're being anecdotal: I almost never want full flattening. My usual use case is that I have an array of arrays of Ts, and would like to have an array of Ts. I'd be very surprised if I sometimes got something else. Infinite depth flattening is really weird if you're modeling the types of your data.

For a concrete use case: suppose I have a list of polygons, where a polygon is represented by a list of vertices given as [x,y]. If I flatten this list, I am expecting to have a list of points, not a list of alternating x coordinates and y coordinates.