Feature request: Code generation for pytorch

[jarlebring]

(Based on a comment from when I presented this result at EPFL.) The generated code is very GPU-friendly. We could generate code also for for PyTorch.

In fact the Bader, Blanes, Casas (in our code graph_bbc) method is currently default in pytorch matrix_exp, and we can do a lot better than the BBC-method so there is certainly room for improvement. It is implemented in the c++ code: mexp_impl and mexp:

https://github.com/pytorch/pytorch/blob/f4dd88489a77ff9b300bf6f9b34c233ca82f76d7/aten/src/ATen/native/LinearAlgebra.cpp#L2111
https://github.com/pytorch/pytorch/blob/f4dd88489a77ff9b300bf6f9b34c233ca82f76d7/aten/src/ATen/native/LinearAlgebra.cpp#L2185
and lower level code compute_T1 ... :
https://github.com/pytorch/pytorch/blob/f4dd88489a77ff9b300bf6f9b34c233ca82f76d7/aten/src/ATen/native/LinearAlgebra.cpp#L1855

I would suggest we add a language with a variant symbol

struct LangCPP
  variant::Symbol
end

If variant=:plain it will generate self-contained C++ code. If variant=:pytorch it generates pytorch-compatible code.