Assistant: reorg/refactor provider handling

[sharon-wang]

We've collected various ideas around cleanup in the provider code, to reduce complexity, improve code quality, and facilitate a better dev experience when working on provider-related code.

• Separate provider classes in models.ts into separate files
• Use more accurate language, e.g. LanguageModel --> ModelProvider
• Make AILanguageModel abstract class more robust
  • make providerName abstract so it must be specified by subclasses
  • try to reuse more code / reduce repeated boilerplate in provider classes
  • can we make it so that the ai client does not necessarily have to be the vercel ai sdk? Some providers may use custom SDKs or the OpenAI SDK
• Consistent logging across all providers, ensure debug/trace/info/warn/error are used appropriately to provide enough logging at the INFO level, which is the default log level
• Add integration tests for each supported provider
• Improve commenting in Assistant extension in general, i.e. JSDoc for functions
• Can we move model filtering utils to Core and expose to extensions via API
• Ensure we error/bail appropriately when models can't be retrieved dynamically and we don't have a default/static list that can be used