fix numerous issues with WeakKeyDict

[vtjnash]

Delay cleanup of WeakKeyDict items until the next insertion. (And fix
get!, since previously usage of it would have added keys without
finalizers to the dict.)

The most significant change here is that I'm changing the way WeakKey works: now instead of the GC clearing the internal reference after the last instance of it disappears from the system (after running finalizers), I'm instead clearing it prior to (happens-before) scheduling any finalization. This is significant because—while this breaks the old implementation of WeakKeyDict—it is a necessary property for precise detection of whether the object is still strongly alive after recovering a reference to it out of a WeakKey object. This happens to be how other languages do it (https://docs.oracle.com/javase/8/docs/api/java/lang/ref/WeakReference.html, https://docs.microsoft.com/en-us/dotnet/standard/garbage-collection/weak-references), so I feel reassured this is the right change.

This also means we could also now drop the internal lock that was need to make this thread-safe (not implemented yet). Thoughts?

Fixes #26939

[JeffBezanson]

I like the general idea but there are still some issues. For example I was able to look up nothing:

julia> d = WeakKeyDict{Any,Any}();

julia> d[[1]]=1
julia> d[[2]]=1
julia> d[[3]]=1
julia> d[[4]]=1
julia> d[[5]]=1

julia> GC.gc()

julia> d[nothing]
1

julia> d
WeakKeyDict{Any, Any} with 5 entries:

julia> 

length is a tough one. It should probably also call _cleanup_locked. Of course it's still hard to use correctly since the true length can change at any time. But we could at least have the property that the length is correct if no GC happens (haha) during the process of getting the length and then iterating.

[vtjnash]

Yeah, I'm not quite sure how to deal with that. I'm considering just defining summary to print "at most", so it's less odd (and define it as SizeUnknown for collect). I don't really want to make length O(n) or mutating, but it is non-trivial then to do better than this.

[Keno]

Overall direction looks right to me.

[JeffBezanson]

It still seems worthwhile to me for length to call _cleanup_locked.