Added array_permutations (attempt 2)

[ronnodas]

Most of the changes are to create a common abstraction that works for permutations and {array_,}combinations, much of it by @phimuemue. Completes a variant of #1001.

[codecov]

Codecov Report

Attention: Patch coverage is 94.17476% with 6 lines in your changes missing coverage. Please review.

Project coverage is 94.38%. Comparing base (6814180) to head (1348785).
Report is 134 commits behind head on master.

Files with missing lines	Patch %	Lines
src/lazy_buffer.rs	89.47%	6 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #1014      +/-   ##
==========================================
- Coverage   94.38%   94.38%   -0.01%     
==========================================
  Files          48       50       +2     
  Lines        6665     6930     +265     
==========================================
+ Hits         6291     6541     +250     
- Misses        374      389      +15     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[phimuemue]

• Importing things: Do we use use something as _ in our codebase? If not, please go with use something;.
• Is ArrayOrVecHelper for Vec<usize> even needed after all this? Shouldn't combinations also run via Box<[usize]>?

[phimuemue]

I think PoolIndex is not really tied to LazyBuffer (it should be compatible with anything you can usize-index into), so it should probably have an own module.

[phimuemue]

Is cycles really an Idx or should it actually be Idx::Item<usize>? I know that these are the same types, but don't the semantics differ? I.e. isn't Idx always meant to index into a slice, whereas Idx::Item<usize> is an array or a Vec containing usizes?

If we'd like to avoid answering this question, we could keep Loaded::k: Idx::Length.

[ronnodas]

cycles is really neither an Idx nor an Idx::Item<usize>, since Idx::Item should be what is obtained by indexing something of type Idx into the buffer, but I don't feel strongly about this. Maybe it makes sense to keep cycles of type Box<[usize]> and switch back to the name PoolIndex so that that trait has clearer semantics.

[phimuemue]

Isn't this indices.extract_item(vals)?

[ronnodas]

No, for the same reason as in the PermutationState::Loaded case, indices is now of length n.

[phimuemue]

Is Idx still a good name if the trait is ArrayOrVecHelper?

[ronnodas]

Is ArrayOrVecHelper for Vec even needed after all this? Shouldn't combinations also run via Box<[usize]>?

This is needed to allow Combinations::reset, which is (only) used by powerset.

[phimuemue]

@jswrenn I'd like to bump to Rust 1.65, because it allows generic associated types.

We currently have trait PoolIndex<T>, but what we actually want is trait ContiguousSequence { type Type<T>: BorrowMut<[T]>; } which could then be impld for Vecs and arrays. Then, we could use ContiguousSequence::Type with different Ts (nice, because the indices/cycles use ContiguousSequence::Type<usize> and e.g. Permutations::Item = ContiguousSequence::Type<UnderlyingIterator::Item>).

I tried staying in Rust 1.63, but we'd have to bend over backwards to achieve something similar.

Please let me know if I can do this, then I'll prepare a PR changing PoolIndex<T> to ContiguousSequence, paving the way for array_permutations.

@ronnodas I'll keep you in the loop.