Fix for OOM problems on GPU at large datasets.

[PhilipMay]

see #522 and #487

TODO

• do testing
• get review & feedback

[PhilipMay]

Internal tests with large data sets show that the code fixed the oom problem.
I made the following changes:

1. I detatch the whole embeddings once instead of each single emb - this should be more efficient
2. If convert_to_numpy I move the whole embeddings to cpu.

IMO this PR is ready for review and merging.

[nreimers]

Thanks for the work. I will have a look today afternoon and merge it