feature: split inference token usage into per-kind counters

[evaline-ju]

Feature Description

Today InferenceExtension carries a coarse TotalTokens / PromptTokens / CompletionTokens triple, plus a partial cache split (CacheReadTokens, CacheWriteTokens). This obscures real cost structure and blocks per-kind policy.

Proposed Solution

• Add a provider-neutral split to InferenceExtension
• Extend session-budget to leverage this

Want to contribute?

• I would like to work on this issue.

Additional Context

No response