I successfully wrote convert_to_cpu for elmo but inference takes time on cpu

With the help of issues and comments reported here on convert_to_cpu.py for ELMO I wrote the script.
I am able to run the ELMO model on CPU, but the converted model is taking time for inference. Also results for GPU and CPU model is slightly different in terms of score for same trained model.

Output of convert to cpu model
```
Rebuilding
These should be close:
[False]
[array([[0.8571025]], dtype=float32)]
[array([[0.84660774]], dtype=float32)]
```