The code that computes the l2 norm of the gradient is going to need
scratch space on every device that DYNET is using that has a parameter.
1) devices should know whether they have parameters/gradients

alignment code is hidden away. it's all hard coded, but it looks like
Intel at least is getting more foregiving about alingment problems so
we might not notice opportunities for speedups if something changes.
GPU memory is aligned mostly by CUDA

the MP stuff needs to be tested by Austin.