Keep track of token usage

[felixarntz]

The AI services infrastructure should keep track of token usage, for both user prompts and model responses. The numbers should be persistently stored (e.g. in user meta or options), so that plugins can use them for reporting or limiting usage of certain features.

Ideally, the token usage would be distributed in distinct buckets per feature - maybe every feature could provide a unique slug to use for this dimension.

[felixarntz]

Potentially helpful library: https://github.com/yethee/tiktoken-php