[V1] [Spec Decode] Add metrics (acceptance rate, number of accepted tokens per step) for V1 rejection sampler

[LiuXiaoxuanPKU]

We need to give a more clear definition of accepted tokens, generated tokens, bonus tokens, and acceptance rate for V1 compared to V0.
We need to track those metrics in the rejection sampler.

[LiuXiaoxuanPKU]

Related files:

1. scheduler:

   vllm/vllm/v1/core/scheduler.py

   Line 105 in 270a5da

   def schedule(self) -> SchedulerOutput:

2. model runner:

   vllm/vllm/v1/worker/gpu_model_runner.py

   Line 917 in 270a5da

   def execute_model(

3. doc: https://docs.google.com/document/d/1xRix4gZqmKAPe6c9mw7B883-vxhlErpQkIBpRgcNZtg/edit?tab=t.0
4. rejection sampler: https://github.com/vllm-project/vllm/blob/main/vllm/v1/sample/rejection_sampler.py

Some questions:

1. accepted tokens, acceptance rate, request level? batch level? workload level?
2. give clear definition of acceptance rate, system efficiency (better name?), reference paper: https://arxiv.org/pdf/2211.17192.
3. Where do we add the metric tracking logic?
4. How to test?