feat: display tokens per second for assistant messages

[edlsh]

Summary

Implements tokens/second display for assistant messages per #5374.

Changes

• Track firstToken timestamp when streaming begins
• Display tok/s rate next to response duration (only for completed text responses)
• Minimum 250ms elapsed time threshold to avoid noisy metrics
• Accumulate tokens instead of overwriting for multi-part responses
• Comprehensive test coverage for token utilities

Notes

If styling is not the direction the team wants to take, I'm happy to make revisions or move the display to a different location.

Closes#5374

[edlsh]

[haltingstate]

Is this merged yet?

[SohaibTaqat]

we need this!!

[edlsh]

No not merged yet. @rekram1-node 😁

[0pilatos0]

Would love to see this merged, +1

[MrHaila]

+1 on possibly getting this merged. Would be a highly interesting metric to follow while evaluating providers and I don't think the added information density hurts usability since there's very little to do while waiting for inference.

[JohnCode-Dev]

Hey @edlsh — nice work on the original implementation here, this was a great reference. Since this one's been sitting with merge conflicts for a while, I went ahead and reimplemented it cleanly on current dev in #12721. Main difference is the TPS calculation lives in core/tokens/ instead of tui/util/ so it's reusable by the SDK and non-TUI consumers.

Full credit for pioneering this — just wanted to get it across the finish line since a few people have been asking for it. If you get a chance, would appreciate your eyes on the new PR!