Add ContextTooLargeError to allow for better context management downstream

[mhpenta]

It would also be useful to provide a ContextTooLargeError, with error translation for anthropic, google, and openai - and compatible.

Why

I am primarily interested in being able to continue conversations in Crush which, for whatever reason, accidentally overflow:

The above is what happens if you overflow + try and compact. To resolve this issue elegantly, we need to understand the error from the API within crush. IMO, the best way tot do that is through error translation. Then crush can recover as it sees fit (perhaps dropping last message which caused the overflow and summarizing).

 // errors.go addition
  type ContextTooLargeError struct {
      UsedTokens int    // Actual tokens in request (if parseable)
      MaxTokens  int    // Model's limit (if parseable)
      Model      string // Model that rejected the request
      Cause      error  // Underlying ProviderError
  }

  // Each provider's toProviderErr() would be modified:
  func toProviderErr(err error) error {
      provErr := // existing logic...

      // New: detect context overflow and wrap
      if ctxErr := parseContextError(provErr); ctxErr != nil {
          return ctxErr
      }
      return provErr
  }

This would require writing translators for at least anthropic (and compat), Gemini and OpenAI (and compat).

I would be happy to create these! Please let me know.