bpart: failed "edge" invalidation

[timholy]

The new tests in #58336 should be made to pass before the release of 1.12.

[Keno]

I think the issue here is that we look at the implicit forward edges from GlobalRefs only for two cases:

• Extext method (new methods added to external generic functions)
• Internal methods (new methods for internal generic functions)

However, in this case, the method is not new (it's in StaleA), but we only store the specialization in StaleB. I think the fix is to use an extra bit in did_scan_source to track whether the method body has been invalidated since jl_require_world.