[Feature]: PD separation supports prefix caching #12257

[skyCreateXian]

Increase support for prefix caching in PD

1. Sending and receiving without transmitting the full amount of KV and hidden signals
2. Resolve the issue of opening prefix caching request errors

[skyCreateXian]

After the prefix hits, mark some of the previous text as context, so only the part that needs to be calculated is transmitted

[ShangmingCai]

After the prefix hits, mark some of the previous text as context, so only the part that needs to be calculated is transmitted

The main branch of this repo will synchronize all updates of the upstream vllm repo, so we recommend that you submit and contribute this PR to the vllm community.