GC may have inconsistent view of stacktop with deferred references #118926

In #118450, Sam pointed out that GC can occur at any Py_DECREF. With deferred references, we will need to traverse the object's stacks. The main problem is that frame->stacktop might not represent the true stacktop at a GC collection, because frame->stacktop is not set in between opcodes.

This requires either scanning the entire stack (not just up to stacktop), or deferring GC to safe points.

Linked PRs

• gh-118926: Deferred reference counting GC changes for free threading #121318
• GH-118926: Better distinguish between pointer and arrays in interpreter generator #121496
• gh-118926: Spill deferred references to stack in cases generator #122748