This avoids copying a new `Providers` struct for each downstream crate
that wants to use it.

Instead, ignore resolution errors that occur in item bodies.

The reason this can't ignore item bodies altogether is because
`const fn` could be used in generic types, for example `[T; f()]`

- Explicitly check for missing docs
- Don't run any lints except those we explicitly specified

- Remove unnecessary `should_loop` variable
- Report errors for trait implementations

These should give resolution errors because they are visible outside the
current scope. Without these errors, rustdoc will give ICEs:

```
thread 'rustc' panicked at 'attempted .def_id() on invalid res: Err', /home/joshua/src/rust/src/libstd/macros.rs:16:9
  15: rustc_hir::def::Res<Id>::def_id
             at /home/joshua/src/rust/src/librustc_hir/def.rs:382
  16: rustdoc::clean::utils::register_res
             at src/librustdoc/clean/utils.rs:627
  17: rustdoc::clean::utils::resolve_type
             at src/librustdoc/clean/utils.rs:587
```

- Add much more extensive tests
  + fn -> impl -> fn
  + fn -> impl -> fn -> macro
  + errors in function parameters
  + errors in trait bounds
  + errors in the type implementing the trait
  + unknown bounds for the type
  + unknown types in function bodies
  + errors generated by macros

- Use explicit state instead of trying to reconstruct it from random info
- Use an enum instead of a boolean
- Add example of ignored error

Instead, report the error.

This emits the errors on-demand, without special-casing `impl Trait`, so
it should catch all ICEs of this kind, including ones that haven't been
found yet.

Since the error is emitted during type-checking there is less info about
the error; see comments in the code for details.

- Add test case for -> impl Trait
- Add test for impl trait with alias
- Move EmitIgnoredResolutionErrors to rustdoc

This makes `fn typeck_item_bodies` public, which is not desired behavior.
That change should be removed once
rust-lang#74070 is merged.

- Don't visit nested closures twice

Previously, rustdoc would issue a delay_span_bug ICE on the following code:

```rust
pub fn a() -> impl Fn() -> u32 {
    || content::doesnt::matter()
}
```

This wasn't picked up earlier because having `type Alias = impl Trait;`
in the same module caused _all closures_ to be typechecked, even if they
wouldn't normally. Additionally, if _any_ error was emitted, no
delay_span_bug would be emitted. So as part of this commit all of the
tests were separated out into different files.

Just use a boolean instead.

Previously two different parts of the context had to be passed
separately; there were two sources of truth.

TyCtxt is a reference type and so can be passed by value.

`evaluate_obligation` can only be run on types that are already valid.
So rustdoc still has to run typeck even though it doesn't care about the
result.

- Move static variables into the innermost scope in which they are used
- Clean up comments
- Remove external_providers; rename local_providers -> providers

This matches the previous behavior of everybody_loops and is also more
consistent than special-casing impls.

This avoids duplicating the same struct twice.