[RFC] V1 engine support

[wangxiyuan]

Currently, ascend works on V0 engine only. This issue tracks the V1 engine support TODO list or known issue:

1. Users must add __main__ check to their script when use V1, see https://docs.vllm.ai/en/latest/design/multiprocessing.html for detail, for example：
   V0:

   from vllm import LLM, SamplingParams
   
   prompts = ["Hello, my name is"]
   sampling_params = SamplingParams(max_tokens=100, temperature=0.0)
   llm = LLM(model="facebook/opt-125m")
   outputs = llm.generate(prompts, sampling_params)
   

   V1

   from vllm import LLM, SamplingParams
   
   if __name__ == "__main__":
       prompts = ["Hello, my name is"]
       sampling_params = SamplingParams(max_tokens=100, temperature=0.0)
       llm = LLM(model="facebook/opt-125m")
       outputs = llm.generate(prompts, sampling_params)
   

2. vllm V1 engin is hard-code to cuda, it need refactor. for example, it uses cudagraph only.
3. V1 runs with torch inductor by default.

   if envs.VLLM_USE_V1 and self.model_config is not None and \
       not self.model_config.enforce_eager:
       # NOTE(woosuk): Currently, we use inductor because the piecewise
       # CUDA graphs do not work properly with the custom CUDA kernels.
       # FIXME(woosuk): Disable inductor to reduce the compilation time
       # and avoid any potential issues with the inductor.
       self.compilation_config.custom_ops = ["none"]
       self.compilation_config.use_cudagraph = True
       self.compilation_config.use_inductor = True
       self.compilation_config.cudagraph_num_of_warmups = 1
       self.compilation_config.pass_config.enable_fusion = False
       self.compilation_config.pass_config.enable_reshape = False
       self.compilation_config.level = CompilationLevel.PIECEWISE
   

4. some new function need be added in vllm-ascend, for example get_kv_cache_spec