[SpecDecode] support spec decode

[MengqingCao]

What this PR does / why we need it?

support spec decode.
plz merge this after vllm-project/vllm#15195 is merged

Does this PR introduce any user-facing change?

N/A

How was this patch tested?

[MengqingCao]

Using this pr will raise segment fault in vllm/spec_decode/metrics.py at

        self._aggregate_num_accepted_tokens = torch.tensor(
            0, dtype=torch.long, device="cpu", pin_memory=pin_memory)
        self._aggregate_num_emitted_tokens = torch.tensor(
            0, dtype=torch.long, device="cpu", pin_memory=pin_memory)

This is a bug form pytorch 2.5.1, reproduction code:

import torch
torch.tensor(0, dtype=torch.long, device="cpu", pin_memory=True)