Make list slightly less embarrassingly slow

[irevoire]

Make list faster in some cases.

Part of #448

On my computer, running:

time cargo run --release -- -e 'sum(map(sqrt, range(0, 10_000)))'

Take 42s on main vs 0.4s on my branch.

I don't think this approach is general enough, though. Having pattern matching would make performant numbat code easier to write.

---

How does it works?

I replaced the NumbatList type with a custom ArcListView type that shares its allocation with all other list referring to the same allocation.
It hold a reference counted ptr to the allocation + the bounds accessible by the allocation:

/// Reference counted list / list view
#[derive(Clone, Eq)]
pub struct ArcListView<T> {
    /// The original alloc shared between all values
    alloc: Arc<VecDeque<T>>,
    /// The indexes accessible to us. If `None` we own the whole allocation
    view: Option<(usize, usize)>,
}

Then, doing a tail is a matter of advancing the starting view by one.
head simply returns the first element.
And push_front check if he's the only owner of the allocation, if it's the case it never allocate (this is always the case while building the list for example so its important).
if it's not the only owner of the allocation, it clones the allocation entirely. We could probably do something smarter here with a chained list or something, but since it yields pretty good results, I didn't try harder.

We now need to take extra care while implementing PartialEq

---

EDIT: While playing with list a little bit I noticed that the following code was still taking around 15s:

let a = now()
let list = range(0, 10_000)
print("Made the list in {human(now() - a)}")

let b = now()
let reversed = reverse(list)
print("Reversed the list in {human(now() - b)}")

Which is expected considering the number of operations we're doing to push something at the end of a list even though we internally uses a vecdeque.
By providing an ffi cons_end function I went from 15s to 0.002s

[sharkdp]

Wow — thank you! Excellent idea and fantastic results! I'm really looking forward to this. I recently played a lot with #73 and that is going to become so much faster.

Can we move ArcListView to a new module (maybe call it just ListView or NumbatList or similar?) and add some unit tests?

I plan to do a more detailed review soon.

[irevoire]

Done ✅
The PR is ready to review again

[sharkdp]

Thank you for the updates and explanations.

Fantastic work!