Consider removing array constants

I stumbled across this page, which lists four constants in the spec: e, inf, nan and pi.

I’m not strongly opposed, but it isn’t really clear to me why we need these in the spec when they already exist in Python’s builtin math built. Given that these should be builtin Python float objects, it isn’t clear to me what we gain from including these, except saving users another import?

[rgommers]

I could live with leaving these out, but personally lean towards leaving them in for convenience. And because I don’t like the math module much, would rather not rely on anything in the stdlib.

[rgommers]

We discussed this in a call today. @kkraus14 had a strong preference for keeping these constants in, because he wants to be able to implement them as a custom Python object that’s a singleton (and duck-types as a float, e.g. by implementing __float__). That way, once it’s on a GPU it can be reused there and this saves implicit syncs, compared to using math.pi et al.

Since no one is strongly opposed to keeping these constants in the API, that’s a good reason to stay with the status quo and keep them.