SIFT memory leak

It seems that the SIFT detection leaks memory. Here is the example script I am running.

```python
import arrayfire as af
import time

while True:
im = af.load_image('image.jpg')
feats, desc = af.vision.sift(im)

af.device.device_gc()
af.print_mem_info()
time.sleep(1)
```

I solved part of the leaks in #168. Apparently the features destructor wasn't called, and af_release_features needed to be called explicitly from python.

With this fix the ORB feature detection works without leaks, but SIFT and GLOH still do.

Running pretty much the same methods in C works without a memory leak. So I think there might be another destructor that ought to be called explicitly, just like af_release_features, but I could not figure out where the leak was originated from.

Start by reproducing the supplied loop with af.load_image, af.vision.sift, af.device.device_gc, and af.print_mem_info, then compare its behavior with the C implementation and the ORB result described in the issue. Trace the feature cleanup around af_release_features and the SIFT/GLOH bindings; done means repeated detection no longer shows memory growth.

Rédigé par le modèle d'indexation à partir du texte de l'issue.