Instrument `count_tokens` / `compute_tokens` methods

## Problem

`Models.count_tokens` and `Models.compute_tokens` (and their async variants) make API calls but are not instrumented. They lack spans for latency and error monitoring.

* `count_tokens` works on both Gemini API and Vertex AI, returns `total_tokens` and `cached_content_token_count`
* `compute_tokens` is Vertex AI only, returns actual token IDs and bytes

## Files

* `sentry_sdk/integrations/google_genai/__init__.py` — `setup_once()` needs additional patches

## Acceptance Criteria

* `count_tokens` and `compute_tokens` calls produce spans
* Async variants are also instrumented
* Tests verify span creation and basic attributes