fix: add dummy_prefill guard for PD connection operations

[FirwoodLin]

Motivation

The current implementation attempts to call shelf_prefill_session on the PD (Prefill-Decode) connection pool regardless of whether dummy prefill mode is enabled. However, when dummy_prefill is set to True, there is no actual prefill session migration happening, and attempting to shelf a non-existent prefill session can lead to unnecessary operations or potential errors.

File "/nvme2/share/root/src/LMDeploy/lmdeploy/serve/proxy/proxy.py", line 643, in chat_completions_v1
    node_manager.pd_connection_pool.shelf_prefill_session((p_url, d_url), prefill_info['id'])
KeyError: 'id'

Modification

Added conditional guards around node_manager.pd_connection_pool.shelf_prefill_session() calls in two API endpoints:

• /v1/chat/completions endpoint [chat_completions_v1]:
  Added if not node_manager.dummy_prefill: guard before calling shelf_prefill_session

• /v1/completions endpoint [completions_v1]: Added if not node_manager.dummy_prefill: guard before calling shelf_prefill_session

Checklist

1. Pre-commit or other linting tools are used to fix the potential lint issues.
2. The modification is covered by complete unit tests. If not, please add more unit tests to ensure the correctness.
3. If the modification has a dependency on downstream projects of a newer version, this PR should be tested with all supported versions of downstream projects.
4. The documentation has been modified accordingly, like docstring or example tutorials.

[grimoire]

LGTM