Guard DP dummy inputs around pending work

[grimoire]

This pr could reduce numbers of dummy inputs when dp>1 is enabled.

[Copilot]

Pull request overview

This PR reduces unnecessary dummy forward inputs when data-parallel (DP) workers have real inputs pending in the CPU-side preprocess queue, improving DP forward efficiency and avoiding wasted dummy steps.

Changes:

• Update DPForwardInputsMaker to treat _pre_in_que as pending real work, waiting briefly for preprocessing to complete before falling back to dummy inputs.
• Add a small “yield to RPC delivery” window to reduce dummy forwards caused by event-loop/RPC scheduling gaps.
• Add targeted unit tests covering the new DP input selection behavior; minor docstring punctuation normalization in guided spec tests.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated no comments.

File	Description
lmdeploy/pytorch/engine/model_agent/inputs_maker.py	Adds pending-real-input detection via _pre_in_que, preprocessing wait, and RPC-yield logic before dummy fallback.
lmdeploy/pytorch/engine/model_agent/agent.py	Clarifies preprocess-loop contract: _pre_in_que represents pending real work until moved to _in_que.
tests/pytorch/engine/test_model_agent.py	Adds unit tests verifying DPForwardInputsMaker.get() behavior across pending, RPC-yield, and sleeping scenarios.
tests/pytorch/spec_decode/test_guided_spec_integration.py	Minor docstring character normalization (ASCII arrow/dash).

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.