[MRV2] Worker-Level Timings

[benchislett]

Purpose

1. With minimal overhead, profile the model runner and speculative decoding proposer (if present) timings
2. Expose the timings via metrics endpoint and also significantly improve the UX and reliability of the iteration-level logging details (see log_worker_iteration_details)
3. Establish reliable statistics for introspective profiling-based tuning such as adaptive verification

[mergify]

Documentation preview: https://vllm--50919.org.readthedocs.build/en/50919/

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @benchislett.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork