This is an initial work in progress for #291.

I have started adding the necessary testing scaffolding to eventually refactor our linear algebra code and, most importantly, drop the accelerate-llvm dependency.

I have done an audit of the codebase and it's quite a lot of work to get rid of accelerate altogether as it's used in a lot of places, however, there is only a single module that uses the llvm backend, which is the Gargantext.Core.Methods.Similarities.Accelerate.Distributional module, which contains the critical function distributional, which is used in the Graph computation algorithm.

The plan of attack is that I have now sufficient infrastructure to start benchmarking distributional. Unfortunately it seems there are no viable backend in accelerate we could use apart from the accellerate-llvm one, so what I will explore is using a different library altogether to implement the distributional function.

A good candidate to explore is massiv, which seems to be a new library which supports multicore and parallelism and it's full Haskell, afaik.

@anoe While I was working on the tests, I have found a bug in the implementation of distributional, which could lead to matrixes populated by NaN. See commit 7a24253fc9eb5668d5752b562e384805ea3d8a5d for more details on the bug.