Support Component aliases and renaming in the inference pipeline

Models may accidentally refer to components that don't exist, by using synonyms or different spellings of existing components.

We can recover from these errors by allowing aliases to be registered in the inference pipeline, to map the names used by the model, to the actual implementations.

Aliases can also be used for prompt tuning. It might turn out that it would be better for `Card` to be called `ContainerWithBorder` from an inference perspective, and that this would improve eval results. We can add a 'renaming' mapping in the inference pipeline so that we can try different names at the inference level, without needing to rename components in our clients etc.