We are now around 6/7 times slower than the LLVM code.

This also changes the cabal.project to not pull the accelerate-llvm or
the llvm-hs dependencies.

There are some rounding errors in the tests, but otherwise
the implementation seems correct.

This commit starts introducing `massiv` in the codebase,
initially for simple functions like `termDivNan`. The main
goal is to extend the linear algebra toolkit up to the
point where we can implement `distributional` in terms of
`massive` and measure its performance.

The previous code was sometimes yielding a matrix of NaN numbers as
it was attempting the division of the input matrix with the diagonal,
which would be 0 in case of an input matrix of 0, resulting in a
division by 0 error.

The main idea is trying to refactor/improve the existing linear algebra
functions one function at the time, using reference implementations and
benchmarks along the way.