GitLab

https://github.com/laekov/fastmoe/issues/63

* Added default weight initializations to FMoELinear and NoisyGate

* Following torch's naming convention

the old impl raised error "too many values to unpack (expected 1)"

* use single variable instead of vector in c functions

* expert count kernel

* remove all lists

* fix old tests

- mask some tensors of tokens for fmoe forward
- pass a list of expert classes to specify what experts in what order want to use