Researching actual token usage

[rajbos]

This pull request enhances the token tracking and reporting features in src/extension.ts by adding support for actual token counts from LLM API usage data, alongside existing text-based estimates. It introduces new fields and interfaces to capture this data, updates aggregation logic to use the most accurate token values available, and improves session log parsing to extract detailed token usage when possible.

Token tracking improvements:

• Added estimatedTokens and actualTokens fields to PeriodStats, session cache, and usage summaries to distinguish between text-based estimates and actual LLM API-reported token counts. [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13]

• Aggregation logic now prefers actual token counts when available, falling back to estimates otherwise, for improved accuracy in usage statistics.

Session log parsing enhancements:

• Introduced PromptTokenDetail and ActualUsage interfaces to capture detailed prompt token breakdowns and actual usage data from LLM API responses.
• Updated session log parsing to reconstruct session state for delta-based formats and extract actual token usage per request, including prompt/completion tokens and prompt token details. [1] [2] [3] [4] [5] [6]

Chat turn details:

• Extended ChatTurn interface and log viewer logic to include actual usage data for each turn, enabling more precise diagnostics and insights. [1] [2] [3]

Cache management:

• Increased cache version to force rebuild and ensure extraction of actual token usage is applied to all sessions.

These changes collectively improve the accuracy and granularity of token usage reporting, supporting both estimated and actual values for better diagnostics and analytics.