GDScript: Respect external refs when cleaning up cyclic refs

[HolonProduction]

Alternative to #114801

Fixes #114420
Fixes #98141 (tested using the 4.3 MRP by myself. Not the original one.)

Supersedes #114623

---

Might be related to: #95428 -> No MRP, but my hypothesis is, that script members were cleared from a still referenced script.

---

When a GDScript is freed it searches for cycles refs amongst its dependencies and clears those as well. However the current implementation does not account for cases were certain scripts are still referenced but by something that is not GDScript. The only indication we have of such a reference is the refcount which the current algorithm does never even look at.

This PR introduces a new approach:

1. Use Tarjan's Algorithm to find strongly connected components (SCC) (almost the same as cycles but we need to account for overlapping cycles) between the transitive dependencies of the script that gets cleaned up.
2. topological sort those.
3. For each SCC count the sum of refcounts, and the sum of internal (from the SCC to the same SCC) references that GDScript knows of. If the difference is 0 the SSC can be cleared as well.
   • If an SCC can be cleared substract it's known references to all other SCC so that those can be cleared as well if they only are alive because of a clearable SCC. -> This works because of the topological sorting.

---

• Pass tests
• Verify linked issues
• Verify that cyclic refs get properly cleaned when it is safe -> tested using this toy project: cycle_cleanup.zip -> requires unreference to actually trigger, that's not an issue with this PR though, GDScriptCache holds references to all scripts so they only get destructed in rare cases
• Cleanup code
  • Find transitive deps during Tarjans DFS instead of doing a second DFS before.

[adamscott]

I would put your new code inside a class TarjansSCCAlgorithm defined in gdscript.cpp. It will make it easier to follow. Also, if possible, it would be nice if the algorithm code symbols to use the same standards as the rest of the code.

Here's a list of names I found:

• strongconnect() parameters and variables:
  • v ➡️ p_visited_node_index
  • w ➡️ visited_node_successor_index
  • head ➡️ stack_head

Also, is this possible to an added boolean for "on the stack/not on the stack" instead of negative index meaning something other than an index?

[akien-mga]

Superseded by #114801.