Issue #3615562: Price cache-write tokens, fix misleading help text
Closes #3615562 https://www.drupal.org/project/ai_metering/issues/3615562
Adds cache_write_per_token to PricingSourceInterface, maps it in LiteLLMPricingSource (cache_creation_input_token_cost) and ModelsDotDevPricingSource (cost.cache_write), and extends TokenEstimator::calculateCost() to charge it. No caller currently supplies a non-zero count -- drupal/ai's TokenUsageDto has no cache-write field to read one from, tracked separately upstream (evidence-based, live-verified against real Anthropic and OpenAI calls: the gap is Anthropic-specific).
Also fixes the misleading cached_tokens Views help text, and makes AiPostGenerateSubscriber::extractTokenDetails() sweep the full TokenUsageDto instead of hardcoding reasoning/total, so a future DTO field is captured automatically.