Version of TopologicalSorter without error checking

[jakelishman]

What is the expected enhancement?

At the moment, TopologicalSorter hashes every index both as it goes through get_ready and as it's returned to done and builds up a hashmap of the size of all visited nodes, only for the purposes of emitting Python-space errors if the user does something incorrectly. None of "putting in an invalid node", "putting in an unready node" and "putting in an already done node" should cause a Rust-space panic (though will ofc cause the algorithm to give nonsensical results).

It'd be nice for performance if there was a way to use a TopologicalSorter that didn't have that memory and runtime overhead. Philosophically, I'd like it if it wasn't there at all, so there's two separate Rust pyclasses, one with error-checking compiled in and one without (and new just builds the correct one), but I think most of my concern would be addressed by making TopologicalSorter::node2state an Option and deciding whether to check it at runtime.

Again, I'll make the PR if there's a chosen direction.