Well, there's only one right now, but it's structured so that we can do more later.

This is the second time I'm implementing this, and the second implementation is much cleaner: to track the error messages, etc., I just use a fine grained enum. That's it! Doing something with string values is a bad idea. :(

Deployment plan for this CL (and any other CLs that gate to experimental):

- [x] wait until a version of crubit that accepts feature flags is released to the stable toolchain.
- [x] change the bzl to implicitly pass in `:experimental` to all targets -- done in 1fd4aec
- [ ] it's now safe to submit gating CLs, including this one.
- [ ] once we're done gating features, apply `:experimental` _manually_ to all targets that need it, and submit a bzl change that reverts #2. (Which targets to apply `:experimental` to, can be determined by the breakage list from the CL reverting #2.)

PiperOrigin-RevId: 676632613
Change-Id: I693be7e380bfefef3db898a9f02512c678148fd9

This is a hacky work around for an unfortunately difficult problem: if they are pub(crate), then Crubit won't know, when looking at a _different_ library, that it needs to write fallback types for it. Instead it just fails. See the newly improved test coverage, ish. (That turned out messier than I hoped.)

After this CL, comprehensive fallbacks for pointers ~works, modulo much smaller gaps, presumably. I'd move my focus onto supporting fallbacks by value.

---

The reasons to make them pub(crate) are:

1. Originally, forward declarations were broken. See b/318690257.

2. visibility restriction -- until forward declarations are generally available, it would be nice to restrict their visibility, so that changing their APIs only breaks O(1) targets.

I fixed the bug in (1) exactly so that it would stop being a reason against this CL. As for #2, going for a cost/benefit calculation, we can just expose them publicly and the blast radius will still not be gigantic. We want to release them as part of `:supported` anyway, so this is just being a little bit aggressive.

The refactor necessary to make these remain `pub(crate)` would have been quite large (it requires every caller to `rs_type_kind` specifying which target they are currently in), and it would only be necessary just to support this one corner case. I think the tradeoff here is worthwhile, let's just do it.

PiperOrigin-RevId: 776235188
Change-Id: I9325bfecb5de0226c9fc5e5e52975864447a4fda