Without this change you need to do:

```
User user;
final SecurityContext oldContext = ACL.impersonate(a);
try {
    user = User.current();
} finally {
    SecurityContextHolder.setContext(oldContext);
}
```

Which seems incredibly wasteful.

Or else you need to replicate the logic in User.current() which makes that code hard to change.

[FIXED JENKINS-20989] PeepholePermalink RunListenerImpl oncompleted should be triggered before downstream builds are triggered

 - queueLengthSnapshot never used.

Ensure that listeners do not get notified when a computer is killed from a thread with the Queue lock

- We don't know what those listeners will be doing
- I'm not completely happy with having to do it this way, there are some theoretical paths where we end
  up re-acquiring the Queue lock in Computer.kill() but they should not be the normal path

Update the retention strategies... the JSR 305 annotations do not seem to be spec'd to apply to child methods also... bold poorly spec'd annotations

- We will never mind that this is probably the wrong import to use as Jenkins core is consistently using the
  wrong one (If anyone asks, for a `javax` import to be correct, it should have a JSR that has passed through voting
  and not have a JSR that is still subject to change... OTOH one could argue that if the JSR 305 should ever
  become revived from its current dormant state *and* it is decided to include in the JRE *then* the packages
  would get moved out of `javax` and into `java` so they are not in final form anyway... but that pre-supposes
  that the process of standardization does not modify the annotations and by virtue of being in `javax`
  there will be some special case handling by some classloaders which is why depending on the JSR305 versions
  is, in my view, a bad plan... anyway this is moot... Jenkins core is going with JSR305 as it currently stands
  so we call all go down together!)

The original method of computing load statistics would compute the total and idle counts independently
which could lead to counting errors while jobs started in between the different state counting operations.
This change switches to returning a `LoadStatisticsSnapshot` so that callers will get a single consistent
view of the counts which was valid for at least one point in time during the collection of the snapshot.