Add integer tensors about marian-dev

[emjotde]

I have now completed a new memory allocator (already in master of marian-train) that will allow to add tensors of different underlying types more easily. All GPU memory chunks are now aligned to 256 byte boundaries which wastes a little memory (insignificant I believe), but apparently makes computations up to 15-20% faster. A nice side effect that isnt't really mentioned anywhere.

I started doing this because I wanted to be able to use the Nervane dot product kernels, as they seemed to be faster than cublas. Turns out this was entirely due to the 256-bytes alignment and now cublas kernels are faster than Nervana.

Typed tensors are next.

[emjotde]

Done, not used much yet, but that will change with time.