Error when using af.set_backend('cuda')

When I try to switch backend using `af.set_backend('cuda')`, I got error like:


It seems that I do not have cuda devices as candiated backend although I have 8 gpus in fact, and the output of `af.info()`
```
In [2]: af.info()
ArrayFire v3.6.2 (CPU, 64-bit Linux, build dc38ef1)
[0] Intel: Intel(R) Xeon(R) CPU E5-2699 v4 @ 2.20GHz, 257588 MB, Max threads(44)
```
and `af.get_backend()`
```
In [2]: af.get_backend()
Out[2]: 'unified'
```
So does I missing something when installing the package?

Start by reproducing the reported af.set_backend('cuda') failure and compare af.info() and af.get_backend() with the reported CPU-only output. Inspect the Python package's backend-loading and installation entry points; done means a correctly installed CUDA-capable environment exposes the CUDA backend or the missing requirement is documented clearly.

Written by the indexing model from the issue text.