improve worst-case performance of BTreeSet intersection v2

[ssomers]

Alternative to pull request #58577: back out of attempts to optimize using ranges, more elegant code (I think).

The stable public type Intersection changes from struct to enum. If that matters, then perhaps changing the fields like in the other proposal also mattered.

[rust-highfive]

r? @joshtriplett

(rust_highfive has picked a reviewer for you, use r? to override)

[Centril]

r? @scottmcm

[Mark-Simulacrum]

@bors try

[bors]

⌛ Trying commit a379504 with merge b75f65a...

[bors]

☀️ Try build successful - checks-travis
Build commit: b75f65a

[scottmcm]

It looks like @KodrAus signed off on the other PR? What's the status of this, given that?

[ssomers]

I have no clue. I don't think any of it was actually merged. I guess tweaking it over and over wasn't the intention but I'm learning. And right now, I think this PR is a better version of the best parts of the first one. It doesn't go as far, but I'm not sure it makes sense to optimize for some cases at some cost for the situations where the previous implementation was best: intersection of small sets. In the "real world" case I'm using this for (through https://github.com/ssomers/Bron-Kerbosch/blob/master/rust/bron_kerbosch/src/util.rs), the further optimization in the other PR doesn't make a difference.

PS better except for the issue that perhaps Intersection being an enum exposes implementation detail.
PPS I found a relevant stackoverflow question and the only answer is basically the way it's done in the other PR (separate struct) but then the author advises against it.
PPPS There are unstable pub enums in the rust code, but I didn't find stable pub enums with unstable variants. And I did find an example of a private enum wrapped by a pub struct: std::ascii::EscapeDefault.