The problem was caused by the improper usage of
`delegate_ctrl` when creating the coreNLP process. For a long
time I was under the impression this flag was essential to allow child
processes to shutdown cleanly without leaving zombie threads, but the
result here in the context of the testsuite was that the coreNLP server
was receiving the first Ctrl^C, completely starving the Haskell RTS,
which wouldn't receive any and as a result our testsuite would be
running forever.

By default, we shouldn't run debug logs for phylo in production, let
alone some that runs within pure code.

Debug logs will hinder performance, and showing them on the production
server is not their place anyway.

This commit correctly propagates the correct `LogConfig` in all the
places where we were just defaulting to log everything, and this allows
us to silence debug logs in tests, unless we want them.

Furthermore, the env var we used to override (in some parts) the logging
level from `LOG_LEVEL`  to `GGTX_LOG_LEVEL`, to avoid the env var
`LOG_LEVEL` clashing with some other service.

This will eventually allow us to properly override the logging level in
the tests, silencing non interesting stuff.

This will allow us to find all the use site of this function, trying to
see if we can use only `withLogger` or correctly control the logging
verbosity via the `HasLogger` and `LogInitParams`.

This commit uses the latest version of `gargantext-graph`, now rebranded
`gargantext-graph-core`, which allowed us to drop unused dependencies
like `accelerate-arithmetic` & co.

The recommended way to build GGTX is via cabal anyway, and stack
shouldn't get in the way of a successful build.

BREAKING CHANGE: this requires a newer version of `nix-shell`, at least
2.18 onwards. You can upgrade following the steps at:

https://nixos.org/download/

In particular, these are the bumped packages:

* GHC (9.4.8 -> 9.6.6)
* Cabal ( -> 3.12.1.0)
* Fortran ( 7 -> pinned)

[worker] various fixes

See merge request !389

Replace performance-critical parts of the algorithm with `massiv` (and ditch `accelerate-llvm`)

See merge request !382

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.

It's crucial to be able to understand the API surface area when porting
the old code to `massiv`. It turns out that other than distributional,
the other piece of code using the LLVM interpreter is
`logDistributional2`.