fix(embedding-linker) optimizer, transformer config, and naming fixes

[adam-sutton-1992]

Hihi,

Three fixes/changes in here for the most part.

1. The embedding linker optimzer doesn't reset upon providing new models. Training would be difficult. So I expose a "load_transformers" to both static and trainable linkers. The trainable model does a bit more with reset_optimizer_and_scaler which points the optimzer to the right model. I also expose learning rate and weight decay, so they're config settings that can be set and changed here.
2. The naming of the embedding linker was inconsistent. So I've set name to trainable_embedding_linker and the class name to Linker. To be consistent with previous design. This was an issue with creating new components and deserializing.
3. With embed_cuis and embed_names being tucked away in the context_model, it didn't make sense to have the model being set there. Somewhat missed from just its static linking days.

[mart-r]

The logged comment looks odd in serialise_to - what it's describing doesn't seem to be happening.

Other than that, looks good.

[adam-sutton-1992]

Fixed that doc issue - from the old embedding linker.