Consider releasing returned values from instance fns

This will collect memory automatically:

(py/run-simple-string "
import torch.utils.data as tud
import torch

class DebugDataset (tud.Dataset):
    def __init__(self):
        pass
    def __len__(self):
        return 1000
    def __getitem__ (self, i):
        return torch.rand(1000, 1000, 1000)

d = DebugDataset()
print(d[0])
print(d[0])")

If the user defines a dataset using clojure and this will cause issues unless the user spams (System/gc):

(let [globals! (py/module-dict (py/add-module "__main__"))]
  (.put globals! "DebugDatasetClj" DebugDataset))

(py/run-simple-string "
d = DebugDatasetClj()
print(d[0])
print(d[0])")

The issue is make-instance-fn runs in the normal jvm context so it is keeping jvm references to the produced items. It does not need to do this; the normal python reference keeping system will work just fine. Fixing this in a robust way will require some thought :-)

Start by tracing make-instance-fn and how it is invoked from py/run-simple-string, then reproduce the dataset examples to observe retained memory. Determine how returned values are held in the normal JVM context and verify that Python reference handling can release them without manual System/gc calls. Done means repeated dataset access no longer accumulates those JVM references.