[WIP] test_keep_alive: Show that keep_alive cycle escapes GC

[EricCousineau-TRI]

This is a PR to track progress for #2761

[YannickJadoul]

I haven't even taken a shallow look, so apologies if I completely missed something, but what about enabling the GC to look into C types, using https://docs.python.org/3/c-api/gcsupport.html ?

[EricCousineau-TRI]

Nice! I don't think it'll fit in the current keep_alive infrastructure as-is, but I believe it could fit in with some sort of indirection (simplest would be cycle-only keep-alive, better would be for any time)

[bmerry]

I haven't even taken a shallow look, so apologies if I completely missed something, but what about enabling the GC to look into C types, using https://docs.python.org/3/c-api/gcsupport.html ?

I think that would be the only sensible way to do it.

[YannickJadoul]

Not sure how, but there's bound to be some kind of way of opting into this, similar to the current py::dynamic_attributes() (which is also translated into a flag set on the actual PyHeapObject, I believe).