DCE bug: unsafe elimination of dead statement, which might not terminate

[aviatesk]

Upon further investigation, it appears that modifying this single line is enough to remove the call to the dead PersistentDict constructor. But, this seems to stem from an inference bug, where we're deleting statements without actually proving terminates.

julia> @noinline Base.@assume_effects :effect_free :nothrow function foo(n)
           s = 0
           while true
               if n - rand(1:10) > 0
                   s += 1
               else
                   break
               end
           end
           return s
       end^C

julia> code_typed((Int,)) do n
           foo(n)
           nothing
       end
1-element Vector{Any}:
 CodeInfo(
1 ─     return Main.nothing
) => Nothing

So, I'd like to address this bug first and then revisit this PR.

Originally posted by @aviatesk in #52954 (comment)

[Keno]

Technically this is not a bug, because we currently still have forward progress guarantees, since we didn't decide on #40009, but I agree, it'd be good to fix this anyway.

[oscardssmith]

I added #40009 to the triage label so we can agree that this is a bug.

[Keno]

Sure. The termination modeling was added specifically to be able to make that change, so we might as well take an official call on it. We'll have to see if we need to improve it though once @aviatesk puts together the change.

[aviatesk]

Submitted a PR: #52999.