Chat: handle transcript length exceeding context window

[abeatrix]

CLOSE https://linear.app/sourcegraph/issue/CODY-1608/chat-token-usage-must-be-updated-before-context CLOSE https://github.com/sourcegraph/cody/issues/4195

• Modify tryAddMessages in PromptBuilder to return undefined when all messages are added successfully
• In DefaultPrompter, throw an error if no transcript was added due to the context window limit being reached
• Add logging to indicate when messages were ignored due to the context limit
• Add a check to throw an error if the token limit is exceeded and no messages were added to the transcript

This ensures the user is aware when the chat input has exceeded the allowed token limit, and the error stops the prompt building process from processing so that it would move to the step where we add context without transcript.

Test plan

1. Build Cody from this branch
2. Pick a model with a smaller context window, e.g. GPT-3.5 Turbo
3. Copy the content from SimpleChatPanelProvider.ts which exceed the context window for the smaller models
4. Submit the question and expect the error about input exceed the context window

[abeatrix]

Anything we can do here?

I was thinking we could truncate the input, but we might truncate the important part of the input without users knowing. I will create a linear issue and get the product for input on what's the best way to handle this case?

Edit: Updated the error message and set it as a transcript message to provide steps for users to unblock themselves: