Faster forward pass + tests on GPU

[Academich]

Hello! Thank you for your CRF implementation! Here is a couple of improvements I would like to propose:

1. Removed the python cycle from the _compute_score method. The same scores can be calculated much faster using pytorch functions without python loops. The difference in speed is especially large on a GPU.
2. Changed the type of the mask from ByteTensor to BoolTensor. If the mask is ByteTensor, pytorch throws warnings that it's deprecated and asks to change the mask type to BoolTensor.
3. Added tests for the forward pass on a GPU. The code of the TestForwardOnCuda class is basically identical to TestForward, but the model and inputs are moved to the GPU first.

[coveralls]

coverage: 100.0%. remained the same
when pulling 966af5a on Academich:master
into 623e340 on kmkurn:master.