add ml_model_settings parameter

[mafrahm]

This PR adds the 'ml_model_settings' parameter that writes key-value pairs directly into the self.parameters attribute.

Example:

class MyModel(MLModel):
    def __init__(
        self,
        *args,
        **kwargs,
    ):
        # we cannot cast to dict on command line, but one can do this by hand
        ml_process_weights = self.parameters.get("ml_process_weights", {"st": 1, "tt": 2})
        if isinstance(ml_process_weight, tuple):
            ml_process_weights = {proc: int(weight) for proc, weight in [s.split(":") for s in ml_process_weights]}

        # store parameters of interest in the ml_model_inst, e.g. via the parameters attribute
        self.parameters = {
            "batchsize": int(self.parameters.get("batchsize", 1024)),
            "layers": tuple(int(layer) for layer in self.parameters.get("layers, (64, 64, 64)),
            "ml_process_weights": ml_process_weights
        }

        # create representation of ml_model_inst
        self.parameters_repr = law.util.create_hash(sorted(self.parameters.items()))

These parameters can then be changed on command line

law run cf.MLTraining --version v1 --ml-model MyModel --ml-model-settings "batchsize=2048,layers=32;32;32,ml_process_weights=st:1;tt:4"

Open questions:

1.) At the moment, the ml_model_settings is not implemented for the MLModelsMixin, therefore we can only use the "default" model (or create new model via derive) when e.g. creating histograms. We could add a ml_models_settings parameter to this mixin, but I'm not sure if this would be used much.
2.) Since the output of the model is now hashed, we might want to also automatically produce an output of the parameters or the parameters_repr such that ml trainings can always be reproduced if necessary.

[pkausw]

Nice PR, thanks! Just some minor questions.

I agree that having a mechanism that dumps the current parameter settings automatically would be nice. However, I think the user can do this themselves at the end of e.g. the training loop. We might want to think about providing e.g. a decorator that can be used directly for this end and that wraps e.g. the training function of the model inst. But imho, this isn't critical for this PR. (Though it also shouldn't be super hard, so if you have some spare time feel free to add it ;) )

[pkausw]

LGTM, thanks!