Plotting stack is a hard dependency: `Plots`, `Cairo`, `PGFPlotsX`, `BenchmarkTools`

`Project.toml` lists all of these in `[deps]`:

```
BenchmarkTools, Cairo, PGFPlotsX, Plots
```

So anyone who wants to mine modal association rules also installs and precompiles a full plotting
stack plus a benchmarking tool. `PGFPlotsX` additionally expects a working LaTeX installation for some
outputs, and `Cairo` pulls binary dependencies.

That is a heavy first install for a user who only wants the rule mining, and it is the kind of cost
that quietly loses people before they reach the interesting part of the package.

### What would close it

Moving the plotting packages behind a package extension (`[weakdeps]` + `[extensions]`), so plotting
lights up when a user loads `Plots` and costs nothing otherwise, and moving `BenchmarkTools` to the
test target.

Julia's extension mechanism makes this fairly mechanical, and the ecosystem already uses it elsewhere -
`SoleModels` has `DecisionTreeExt` and `XGBoostExt`.

Reported from an outside survey of the ecosystem; happy to prepare a pull request if that would help.