support release pipeline

[irexyc]

Motivation

#3055
#2498

Currently only turbomind backend is supported and ytorch backend support should be added after #3051

from lmdeploy import pipeline
pipe = pipeline('/nvme/shared/vicuna-7b-v1.5/')
pipe('hello')
pipe.close()

# for pytorch backend & vl models, you need to invoke below functions manually
import gc
gc.collect()
torch.cuda.empty_cache()

from lmdeploy import pipeline
with pipeline('/nvme/shared/vicuna-7b-v1.5/') as pipe:
    pipe('hello')

# for pytorch backend & vl models, you need to invoke below functions manually
import gc
gc.collect()
torch.cuda.empty_cache()

There are two other issues that do not affect the usage

1: pipe.close will cancel all async tasks, but can't remove the callback. The callback in internal thread will raise error like below, while this will not blobk main thread.

exception calling callback for <Future at 0x7fe1541b6080 state=cancelled>
Traceback (most recent call last):
  File "/home/chenxin/miniconda3/envs/310/lib/python3.10/concurrent/futures/_base.py", line 342, in _invoke_callbacks
    callback(self)
  File "/home/chenxin/ws/lmdeploy/lmdeploy/serve/async_engine.py", line 440, in <lambda>
    asyncio.run_coroutine_threadsafe(_infer(), loop).add_done_callback(lambda x: x.result())
  File "/home/chenxin/miniconda3/envs/310/lib/python3.10/concurrent/futures/_base.py", line 449, in result
    raise CancelledError()
concurrent.futures._base.CancelledError

2. pipe.close() won't clear all cuda resources. But if we repeatedly create and close the pipeline, the cuda memory will not exceed a small value.

[lvhan028]

@zhulinJulia24 may add TC for this PR