Offload safe mutable args with Region and PartitioningStrategy

[Sa4dUs]

Now Region are lang items and mapped as slices.
needs #156620 to work

r? @ZuseZ4

[ZuseZ4]

Next round. I think we should actually refuse all cases (for now) where we have a Region within any ADT/struct/slice, etc. not just &Region. Most of them would be transfered incorrectly, and I don't think we gain much. The other direction is something like Region<Vec<f32>, S>, where we also transfer T incorrectly. But to be fair, I think that direction is already unchecked on main. I assume that offloading a vec will still do the wrong thing, so I'm fine if we accept that here as well till we implement a proper handling. But also feel free to already reject it already if you want.

[rustbot]

This PR was rebased onto a different main commit. Here's a range-diff highlighting what actually changed.

Rebasing is a normal part of keeping PRs up to date, so no action is needed—this note is just to help reviewers.

[ZuseZ4]

I left another round, mostly about comments and tightening the checks.
I feel like we'll need to restrict get and get_mut a bit further (or at least describe them better), but we can leave that for the last round (and discuss on Wednesday).

Now that it seems closer to the final version, can you add a description to
https://rustc-dev-guide.rust-lang.org/offload/internals.html ?
Probably best as a new subsection, so that we have one about the build pipeline, and one about the safety/region design?

View changes since this review

[ZuseZ4]

Is there a reason why you added another feature instead of reusing the existing one?

[ZuseZ4]

same here, can you reuse gpu_offload as fueatre?

[ZuseZ4]

Can you make it more explicit (here and in the trait safety note) that this is also soundness-critical?
In summary, this is only allowed to unconditionally return Ok iff. the given strategy can run with arbitrary len+grid+block values. If there are any combinations where an impl would return incorrect results or cause UB, then check_launch must return a LaunchError. We can later ask Ralf or someone unbiased to check that our description is unambiguous.

[ZuseZ4]

Ty::walk skips a few relevant cases:

     2     /// Iterator that walks `self` and any types reachable from
     1     /// `self`, in depth-first order. Note that just walks the types
  2175     /// that appear in `self`, it does not descend into the fields of
     1     /// structs or variants. For example:

E.g.

struct Wrapper<'a> {
    region: Region<'a, f32, Dummy>,
    scalar: f32,
}

[ZuseZ4]

It seems like I can skip this this test safely by just replacing

offload! {... args = (foo,bar), }

with

let t = (foo,bar);
offload! { ... args = t, }

Just to get started, you can just reject the second case and leave a fixme note in a test and macro for now, we can do something smarter later.