Reduce number of allocations done by NLL

A couple of easy wins. Here are the NLL speedups that exceed 1%:
```
sentry-cli-check
        avg: -3.5%      min: -3.5%      max: -3.5%
inflate-check
        avg: -1.9%      min: -1.9%      max: -1.9%
inflate
        avg: -1.7%      min: -1.7%      max: -1.7%
clap-rs-check
        avg: -1.6%      min: -1.6%      max: -1.6%
cargo-check
        avg: -1.6%      min: -1.6%      max: -1.6%
ripgrep-check
        avg: -1.4%      min: -1.4%      max: -1.4%
serde-check
        avg: -1.2%      min: -1.2%      max: -1.2%
regex-check
        avg: -1.0%      min: -1.0%      max: -1.0%
sentry-cli
        avg: -1.0%      min: -1.0%      max: -1.0%
```
r? @nikomatsakis

Update the error message for a missing global allocator

Don’t mention `#[default_lib_allocator]` (which is an implementation detail irrelevant to most users) and  instead suggest using `#[global_allocator]`, which is often the correct fix.

NLL diagnostics: revise `fn check_access_permissions`

NLL: revise `fn check_access_permissions` so that its (still branchy) shares more code paths between the different cases, and also provide more diagnostics in more cases (though the added diagnostics still do not always meet the quality bar established by AST-borrowck)

----

Transcribing "checklist" suggested by Niko, except I am rendering it as a table to make it clear that I do not regard every item in the list to be a "must have" for landing this PR.

goal | does this PR do it?
-----|------------------------------
no suggestions for `ref mut` |  yes
suggestions for direct local assignment (`{ let x = 3; x = 4; }`) | yes (see commits at end)
suggestions for direct field assignment (`{ let x = (3, 4); x.0 = 5; }` | yes (see commits at end)
suggestions for upvars (`let x = 3; let c = \|\| { &mut x; }`) | yes

Note that I added support for a couple of rows via changes that are not strictly part of `fn check_access_permissions`. If desired I can remove those commits from this PR and leave them for a later PR.

Fix #51031
Fix #51032
(bug #51191 needs a little more investigation before closing.)
Fix #51578

Rename OOM to allocation error

The acronym is not descriptive unless one has seen it before.

* Rename the `oom` function to `handle_alloc_error`. It was **stabilized in 1.28**, so if we do this at all we need to land it this cycle.
* Rename `set_oom_hook` to `set_alloc_error_hook`
* Rename `take_oom_hook` to `take_alloc_error_hook`

Bikeshed: `on` v.s. `for`, `alloc` v.s. `allocator`, `error` v.s. `failure`

Don’t mention `#[default_lib_allocator]` (which is an implementation detail
irrelevant to most users) and instead suggest using `#[global_allocator]`,
which is often the correct fix.

(since I made this mistake at first but the tests didn't catch it):
we should not suggest adding `mut` to a reassigned `ref` or `ref mut` binding.

(The Rust language, since at least 1.0, does not have `mut ref mut` or
`ref mut mut` etc.)

(Follow-on commits updating the test suite show the resulting changes
to diagnostic output.)

Tried to unify various common code paths and also vaguely approximate
the AST-borrowck diagnostics.

The change in (subjective) quality of diagnostics is not a universal
improvement. But I think this is a better code base to work from
for future fixes.

As a drive-by, removed some dead-code.

Namely, we thread down the `HirId` of the explicit type of the
argument.  In the case of the special `self` variable with an implicit
type, we also thread down a description of its structure (`self`/`mut
self`/`&self`/`&mut self`).

Thread info about form of variable bindings, including spans of arg types, down into `mir::LocalDecls`.

As a drive-by: the ref_for_guards created by `fn declare_binding`
should not have been tagged as user_variables in the first
place. These secret internal locals are *pointers* to user variables,
but themselves are not such (IMO. For now at least.)

Use quiet tests by default

r? @EddyB

Use String, not &str in some collection examples

Discussed in #46966

Overuse of borrowed values in data structures is a common mistake I see in Rust user forums. Users who copy&paste such examples end up fighting with the borrow checker as soon as they replace string literals with some real values.

This changes a couple of examples to use `String`, and it adds opportunity to demonstrate use of `Borrow`.

Update books for next release

Add ability to apply custom derive to union types.

The Union item type has been included in the allowed types for a custom derive.
fyi @abonander

Closes #50223

Run some stuff in parallel

Requires https://github.com/rust-lang/rust/pull/50699 to actually work correctly.

r? @nikomatsakis

Fix processing mod with multi-level path on Windows

Fix error in [rustfmt](https://github.com/rust-lang-nursery/rustfmt/issues/1754) because libsyntax can not handle `mod` with multilevel path on Windows.

Alternative is do almost same in https://github.com/rust-lang/rust/blob/master/src/libstd/sys/windows/fs.rs#L717 to allow work on paths with different separators, Ex. "\\\\?\\c:\\windows/temp"

The Union item type has been included in the allowed types for a custom
derive. Closes #50223

These vectors are always small, so this avoids lots of allocations.

`has_any_child_of` is hot. It allocates a `Vec` that almost always
doesn't exceed a length of 1.

This patch peels off the first iteration of the loop, avoiding the need
for the `Vec` creation in ~99% of cases.

Declare DebruijnIndex via newtype_index macro

Part of #49887

Declare `DebruijnIndex` via the `newtype_index` macro.

The acronym is not descriptive unless one has seen it before.

* Rename the `oom` function to `handle_alloc_error`. It was **stabilized in 1.28**, so if we do this at all we need to land it this cycle.
* Rename `set_oom_hook` to `set_alloc_error_hook`
* Rename `take_oom_hook` to `take_alloc_error_hook`

Bikeshed: `alloc` v.s. `allocator`, `error` v.s. `failure`

Stabilize std::path::Path::ancestors

Closes #48581

r? @BurntSushi

Improve memoization and refactor NLL type check

I have a big branch that is refactoring NLL type check with the goal of introducing canonicalization-based memoization for all of the operations it does. This PR contains an initial prefix of that branch which, I believe, stands alone. It does introduce a few smaller optimizations of its own:

- Skip operations that are trivially a no-op
- Cache the results of the dropck-outlives computations done by liveness
- Skip resetting unifications if nothing changed

r? @pnkfelix

Add existential type definitions

Note: this does not allow creating named existential types, it just desugars `impl Trait` to a less (but still very) hacky version of actual `existential type` items.

r? @nikomatsakis