Adding new variables breaks inference

Whenever we add new inputs or outputs to our models, there is a high chance that we break inference because some components of our inference pipeline (namely the metadata patches and the GRIB templates) are hardcoded with a limited set of variables. This forces us to merge hotfixes such as #128, which is not ideal.

We should:

* change the GRIB templating approach so that we don't need to specify the list of params
* create variables metadata patches starting from the anemoi datasets used for training (the current patches were derived from some checkpoints), which contain all the possible variables we could need in inference

Start by tracing the inference pipeline's GRIB templates and metadata patches, then inspect the anemoi datasets used for training. Confirm how the current hardcoded variable lists are derived and compare them with the datasets' available variables. Done means new model inputs and outputs no longer require manual parameter-list updates or hotfix metadata patches.

Written by the indexing model from the issue text.