Fix: Graceful Degradation to CPU if CUDA isn't available

[KernAlan]

Issue

#29

Solution

Include a check using torch if CUDA is available. If it's not, gracefully degrade to cpu.

Testing

Config json set to auto. But the torch cuda check fails, and the model gracefully degrades to cpu:

[savbell]

Thanks for the PR! :)

One request: Can you add back the compute_type parameter when running on the CPU? That way, someone can set it to float32 to avoid the warning that comes up (more info: https://opennmt.net/CTranslate2/quantization.html). Thanks!

[KernAlan]

Done!