fix: make token counting API requests asynchronous for Gemini/Anthropic

[roomote]

This PR resolves #3666 by making the token counting API requests asynchronous, so they no longer block the main inference request.

Changes

• Modified countTokens() in GeminiHandler to return tiktoken estimate immediately while API call happens in background
• Modified countTokens() in AnthropicHandler to return tiktoken estimate immediately while API call happens in background
• Added countTokensAsync() private method to both providers to handle the actual API call
• Added tests to verify asynchronous behavior and immediate returns
• Vertex provider automatically inherits the fix from GeminiHandler

Benefits

• Main inference requests start immediately without waiting for token counting
• Preserves accurate token counts from the API (they just arrive asynchronously)
• Fallback to tiktoken ensures we always have a working estimate
• No breaking changes to the API

Testing

• All existing tests pass
• Added new tests specifically for the asynchronous behavior
• Verified that token counting returns immediately (< 100ms) even when API call takes longer

Fixes #3666

---

Important

Make token counting asynchronous in GeminiHandler and AnthropicHandler, improving performance by not blocking main requests.

• Behavior:
  • countTokens() in GeminiHandler and AnthropicHandler now returns tiktoken estimate immediately, with API call in background.
  • Introduces countTokensAsync() in both handlers for asynchronous API calls.
  • Vertex provider inherits changes from GeminiHandler.
• Testing:
  • Adds tests in anthropic.spec.ts and gemini.spec.ts to verify asynchronous behavior and immediate returns.
  • Ensures token counting returns immediately (< 100ms) even if API call is delayed.
• Misc:
  • No breaking changes to the API.
  • Logs errors to console if async API call fails.

^{This description was created by for 0f08bce. You can customize this summary. It will automatically update as commits are pushed.}

[daniel-lxs]

Incorrect approach, the issue isn't scoped with enough detail