Refactor and document training params (TBA)

Currently, the params data model has a pretty good structure. However, there are things that can be improved. The list of such things is provided below and is being edited from time to time. Along with that, we should start documenting the params. Then it should become more clear what should be changed.

- `templates.py` --> `cl_default_config` `lm_default_config` ?
- package `params` -> `config` !
- `ClassifierTrainingParams` -> `ClassifierConfig`
- `LangModelTrainingParams` -> `LangModelConfig` or `LMConfig`
- `backwards` from `lang_model_training` to `data`
- remove `min_freq` param
- classifier does not use `clip` from params

Start by mapping the current params model across templates.py, the params package, lang_model_training, and data. The issue lists possible renames, moves, removals, and documentation work, but the intended final structure is still to be clarified. Done would require an agreed scope, the corresponding refactor, and documentation of the resulting training parameters.

Written by the indexing model from the issue text.