Fixing those might bring us forward on the path to Numpy 2.0
compatibility.

The recommended way is not to create a random generator instance (e.g.
via `np.random.default_rng()` and use that.  It is supposed to be
significantly faster and statistically more robust.

In case we ever want to be able to set a seed for cluster_utils itself,
it's probably best to have one global generator used throughout the
package.  Add this, together with a little getter function `get_rng()`
in base/utils.py.

This fixes ruffs NPY warnings.

Show a more informative error message when the parameter dict, passed to
the job script, cannot be parsed by `ast.literal_eval`.  It now includes
the string to be parsed itself, to make debugging easier.

Since the string, that is passed here, is constructed by cluster_utils
when used normally, an error here most likely indicates a bug in
cluster_utils rather than a user mistake.

Not too long ago, NumPy 2.0 got released.  cluster_utils didn't work
with it out of the box, so we restricted the dependency to numpy<2.

Looking into it now, it seems the only actual problem was in the
`Discrete` distribution when using booleans:  Using NumPy's `choice()`
converted the list of native bools to an array of np.bool.  Later one,
this clashed when passing the parameters to the job script, where they
are parsed with `ast.literal_eval()`.
Simply converting back to list with `.tolist()` seems to be enough to
fix this.