Use arraymancer for tensor, matrix, and vector

[neonnnnn]

In v0.1.0 nimfm uses its own data type Tensor, Matrix, and Vector (defined in tensor.nim) for parameters and some caches.
It would be good to replace them wtih arraymancer tensor, and to develope fast implementation for dense datasets by using arraymancer (it provides cuda support)

TODO

• Replace tensor.nim with arraymancer
• Develop arraymancer-based-implementation of FMs for dense dataset