feat: collect token usage

[mdjastrzebski]

Summary

Collect token usage during all types of LLMs requests.

Details

1. Added usage field to BaseWorkflowState type, of the CompletionUsage type imported from OpenAI SDK.
2. Modified all LLMs calls to also return usage stats

Alternatively

Alternatively instead of having a separate accumulator variable on state, it could be calculated from messeges if there were to hold it as well. There would also be a need to capture the orchestration usage (e.g. selectAgent) as these are not directly reflected in the messages.

Test plan

Tested manually on surprise_trip.ts example

Resolves #8

[grabbou]

Thanks for the PR! I've been looking into this and potentially thinking about the second approach

Alternatively instead of having a separate accumulator variable on state, it could be calculated from messeges if there were to hold it as well. There would also be a need to capture the orchestration usage (e.g. selectAgent) as these are not directly reflected in the messages.

There's pending rewrite of the internals to enable more customization, partially inspired by this PR. Once that is complete and the PR is up, we will revisit this one (likely tomorrow, given our ambitious plans)

[mdjastrzebski]

In order to handle non-message workflow steps, you could collect the results of processing steps in some thing like audit-log of all workflow actions and then calculate total tokens from there