fix: consume duplicates in rolling_by window

[ritchie46]

@MarcoGorelli FYI

[MarcoGorelli]

nice! so, salvaging the fastpath rather than conditionally taking a slowpath

Just as a suggestion, is it possible to keep the parametric tests from #11237? They're very fast to run, and would add quite some confidence in the method's correctness (it's from them that I discovered the duplicates bug)

[ritchie46]

Sure, can you open them as a separate PR?

[MarcoGorelli]

#11262

good news is - those tests currently fail, but when rebased onto this branch, they pass! very confident in group_by_rolling's correctness then 🎉

[ritchie46]

so, salvaging the fastpath

Yes, I didn't like a full data pass to check if we have duplicates. This shouldn't have any noticable performance implications on the happy path, most of our time is in the binary search and the Duration::add_us.

I think we should take a look at both later.