13 proper handling of batches

[pzimbrod]

Resolves #13

[pzimbrod]

Following things had to be changed:

• For sake of extensibility, NNlib's batched routines have been ditched by OMEinsum. Looping over 3-dimensional arrays would suffice for a 1D Problem, but certainly not for higher-dimensional ones.
• Pre-allocation of arrays had to be dropped for now. In combination with OMEinsum, gains were marginal at best anyway.
• Fixed a problem where the constructor and subsequent training wouldn't work when bias was set to false.
• Typed every part of the data structure. This should allow the LLVM Compiler to produce more optimized code.
• Ditched the batchsize argument out of the initialization function as it's no longer needed. This increases the flexibility of the layer considerably.
• Had to do some permutation at the beginning and end of the layer pass. Otherwise it's not possible to satisfy the sequential requirement of batch dims that CuFFT poses and the requirements of Flux.DataLoader at the same time.