Backend choice reset inside of threads

Inside a new thread, Arrayfire uses the default backend rather than the backend set by the parent thread:

```python
>>> import arrayfire as af
>>> af.get_available_backends()
('cpu', 'cuda', 'opencl')
>>> af.get_active_backend()
'cuda'
>>> af.set_backend('opencl')
>>> af.get_active_backend()
'opencl'
>>> from threading import Thread
>>> Thread(target=lambda: print(af.get_active_backend())).start()
cuda
```

Is this expected behavior and/or documented anywhere? I would think that the active backend should be process local rather than thread local.

If it matters, here's my version info:
OS: Ubuntu 20.04
arrayfire: 3.9.0 (commit [955152b6](https://github.com/arrayfire/arrayfire/commit/955152b6570c608ae74ebd9e6b31d48351cb8a16))
arrayfire-python: 3.8.10

cuda: 11.4.0-1
ocl-icd-libopencl1: 2.2.11-1
intel-mkl: 2020.0.166-1