Be more careful about iterator invalidation during recursive invalida… (JuliaLang#57934)

…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 JuliaLang#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 JuliaLang#57696

Author: Keno

test/invalidation.jl

@@ -283,3 +283,8 @@ begin take!(GLOBAL_BUFFER)
     @test isnothing(pr48932_caller_inlined(42))
     @test "42" == String(take!(GLOBAL_BUFFER))
 end
+
+# Issue #57696
+# This test checks for invalidation of recursive backedges. However, unfortunately, the original failure
+# manifestation was an unreliable segfault or an assertion failure, so we don't have a more compact test.
+@test success(`$(Base.julia_cmd()) -e 'Base.typejoin(x, ::Type) = 0; exit()'`)