### What this PR does / why we need it?
Fix e2e data parallel test: add resource release code and give more time
to engine to pause their processing loops before exiting.
### Does this PR introduce _any_ user-facing change?
No
- vLLM version: v0.9.2
- vLLM main:
5895afd780
Signed-off-by: leo-pony <nengjunma@outlook.com>
### What this PR does / why we need it?
Add `max_num_tokens_across_dp` to AscendMetadata to fix dp
This pr fixes the bug introduced by
https://github.com/vllm-project/vllm-ascend/pull/1229, which add an arg
`max_num_tokens_across_dp` when dp_size > 1.
Signed-off-by: MengqingCao <cmq0113@163.com>