Upgrade to rkyv 0.8

[ThierryBerger]

• ⚠️ To be noted that data serialized by rkyv 0.8 is not compatible with data serialized with rkyv 0.7.

Helpful resources

• rkyv talk: https://www.youtube.com/watch?v=ON4z2LbTD-4
• glam migration: Update rkyv to latest 0.8 release bitshifter/glam-rs#605
• wasmer migration: Upgrade rkyv wasmerio/wasmer#5142

helpful commits from rkyv

• rkyv validation change: rkyv/rkyv@172b451#diff-f2e5d2d70a5fc0728c12e8f58c7f54f5200b7c98cd09cedd89fb55c07c2b711e
• rkyv size_32 to pointer_width: rkyv/rkyv@d592789#diff-f2e5d2d70a5fc0728c12e8f58c7f54f5200b7c98cd09cedd89fb55c07c2b711e

Current state

• deny(missing_docs) fails because of Add documentation for struct fields rkyv/rkyv#603
• compiling, a few tests have been commented and are in progress of being re-introduced.
• I'm exploring rkyv 0.8 capabilities in a "minimum codebase" at https://github.com/vrixyz/rkyv-playground

Difficulties

• I did not succeed in keeping rkyv(as = ) for Matrix.
  • I thought that was because i32 is not Portable. See WIP migration to 0.8 rkyv/rkyv_contrib#4 for a somewhat similar situation.
  • But I was able to implement it for ArrayStorage, so it must be something else in play.
  • I believe using as = makes it impossible to use rkyv(derive(...)), so we have no other choice to implement those manually. It's a bit boiler-plate-y because of _le rkyv types.

Previous analysis for `as`

In current nalgebra, we're using:

#[repr(transparent)]
#[derive(Copy, Clone, PartialEq, Eq, Hash)]
#[derive(rkyv::Archive, rkyv::Serialize, rkyv::Deserialize, bytecheck::CheckBytes)]
#[archive(
    as = "ArrayStorage<T::Archived, R, C>",
    bound(archive = "
        T: rkyv::Archive,
        [[T; R]; C]: rkyv::Archive<Archived = [[T::Archived; R]; C]>
    ")
)]
pub struct ArrayStorage<T, const R: usize, const C: usize>(pub [[T; R]; C]);

I believe we'll want to keep using that as = ; I'm not sure what's the advantage of that exactly, I guess it limits the number of types, to avoid ending up with too many archived types, as well as avoid type juggling. (so compilation and runtime performance) ; ❓ any guess confirmation is welcome 😄

I ended up with this migration, mostly reading (and forcing) through errors:

#[repr(transparent)]
#[derive(rkyv::Archive, rkyv::Portable, Serialize, Deserialize, bytecheck::CheckBytes)]
#[rkyv(
        as = ArrayStorage<T::Archived, R, C>,
        archive_bounds(
            T: rkyv::Archive,
            [[T; R]; C]: rkyv::Archive<Archived = [[T::Archived; R]; C]>
        )
    )
]
pub struct ArrayStorage<T, const R: usize, const C: usize>(pub [[T; R]; C]);

• Implementation of PartialEq between arraystorage is currently not generic

[ThierryBerger]

The debug check underneath is adding a constraint on Archived to be Debug, which was not explicit before, I think that's what we'll want to do like what I did for Matrix through archive_bounds(<S as Archive>::Archived: core::fmt::Debug), but it feels like it's adding potentially unnecessary constraints.

Could a manual implementation could implement Debug only for compatible types ?

Also, the debug format test is failing currently.

[ThierryBerger]

going through ~as = Matrix<Archived::T> and leveraging Portable is probably the best path forward, but I'm not sure how exactly.

[ThierryBerger]

I feel like this is a bit constraining ? This is only needed for automated derive (needed for tests)

See https://github.com/dimforge/nalgebra/pull/1502/files#r2025182401

[ThierryBerger]

Do we want to derive rkyv::Portable ? I'm not sure what the implications are.

I think this would allow us to avoid a specific ArchivedMatrix being created, which is likely what we want, as was the case with as Matrix.

[ThierryBerger]

This as being removed is a UX downgrade because we can't access underlying functions from original type on rkyv archived type.

[ThierryBerger]

We should add a new test on i32 or something different than f32, as currently this would fail due to missing partialEq from arraystorage.