Be more careful about iterator invalidation during recursive invalida…

[Keno]

…tion

It is possible for one MethodInstance to have a backedge to itself (typejoin is a prime example). This didn't used to be much of a problem prior to #57625 because we would just delete the backedges list at the top of invalidation. However, now that we're more selective, we need to be careful not to move backedges around while a frame higher on the stack may be looking at it. To this end, move the compaction part of the deletion into a separate pass and only delete (but don't move around) backedges while a frame higher on the stack may be looking at it.

Fixes #57696

[timholy]

Since I see this is queued for backporting, it's worth mentioning that this broke SnoopCompile (the in-progress update for 1.12).
Specifically, in https://github.com/timholy/SnoopCompile.jl/blob/469db0fb77249bf406f5563d5c7da1586d8ca052/test/snoop_invalidations_new.jl#L160-L164 the logging stream is missing

MethodInstance for Main.MethodLogs.f(::Integer)
  "jl_method_table_insert"
  MethodInstance for Main.MethodLogs.f(::Signed)
  "jl_method_table_insert"
  f(::Signed) @ Main /tmp/SnoopC/snoop_invalidations_new.jl:165
  "jl_method_table_insert"

[timholy]

Here's a MWE. The distinguishing characteristic is the presence of both a poorly-specialized and invoked caller with the same signature:

f(::Integer) = 1
callsfrts(x) = f(Base.inferencebarrier(x)::Signed)
invokesfs(x) = invoke(f, Tuple{Signed}, x)

# compilation
invokesfs(1)                        # invoked callee
callsfrts(1)                        # runtime-dispatched callee

# invalidation
logmeths = ccall(:jl_debug_method_invalidation, Any, (Cint,), 1);
f(::Int) = 2
f(::Signed) = 4
ccall(:jl_debug_method_invalidation, Any, (Cint,), 0);

m = which(f, (Signed,))
m ∈ logmeths

You want the final line to be true.

Here's the backedge graph before invalidation:

julia> m = only(methods(f))
f(::Integer)
     @ Main REPL[1]:1

julia> mis = collect(Base.specializations(m))
2-element Vector{Any}:
 MethodInstance for f(::Int64)
 MethodInstance for f(::Signed)

julia> mi = mis[1]
MethodInstance for f(::Int64)

julia> mi.backedges
2-element Vector{Any}:
 Tuple{typeof(f), Signed}
 CodeInstance for MethodInstance for invokesfs(::Int64)

julia> mi2 = mis[2]
MethodInstance for f(::Signed)

julia> mi2.backedges
1-element Vector{Any}:
 CodeInstance for MethodInstance for callsfrts(::Int64)

[timholy]

Filed as #58080 so it doesn't get lost

[KristofferC]

This is already backported in #57972.