Consolidate yaml schema and configs

[eu9ene]

Currently, when adding/changing a new setting in a Taskcluster experiment config we have to update it in multiple places:

• train action schema
• CI default parameters schema
• CI default parameters values
• Reference production config in taskcluster/configs
• CI yaml config in taskcluster/configs that we currently don't use

We should consolidate to :

• production reference YAML config (same as now)
• CI YAML config in taskcluster/configs instead of the one in parameters json
• one YAML schema that's used for validation in train action and elsewhere

[eu9ene]

Also, I'm still trying to figure out what taskcluster/test/params/large-lt-en.yml and taskcluster/test/params/small-lt-en.yml are for. It seems those are required for some tests and also need to be updated. Not updating them breaks task taskgraph-diff. Anyway, the tests should also use the reference production/CI configs from taskcluster/configs.

[bhearsum]

Also, I'm still trying to figure out what taskcluster/test/params/large-lt-en.yml and taskcluster/test/params/small-lt-en.yml are for. It seems those are required for some tests and also need to be updated. Not updating them breaks task taskgraph-diff. Anyway, the tests should also use the reference production/CI configs from taskcluster/configs.

Yes - these are used to generate graphs with and without some changes applied, and generate a useful to see how a code change affects graphs.

I agree that this could probably be reworked to pull in at least some things from a separate place. One of the advantages of having these concrete files, though, is that it allows us to have multiple versions. At the moment, we just have two with more and fewer datasets, but we could have variants with and without opuscleaner/opustrainer, with and without publication, with various training continuation configurations, etc.