fix: parse multimodal tool messages

[CUHKSZzxy]

Summary

• Parse multimodal content in tool messages instead of only user messages.
• Preserve tool-message metadata such as tool_call_id while converting multimodal parts.
• Add regression coverage for a tool-result image payload and multimodal input detection.

Validation

• Diff whitespace check passed.
• Syntax compilation passed for touched Python files.
• Replayed the original multimodal tool-result chat-completions payload against a real Qwen3.5 LMDeploy server and confirmed it completes normally instead of returning a prompt-processing error.

Assistance

Assisted with Codex + GPT-5.5 xHigh Fast

[Copilot]

Pull request overview

This PR updates LMDeploy’s multimodal message preprocessing so that multimodal tool messages (e.g., tool results containing image parts) are parsed the same way as multimodal user messages, while preserving tool-message metadata needed for tool-call correlation.

Changes:

• Extend MultimodalProcessor._parse_multimodal_item to parse multimodal tool messages (not just user) and preserve message metadata when rewriting content.
• Refactor multimodal-input detection to use a shared MULTIMODAL_TYPES constant and broaden detection to include tool messages.
• Add regression tests for tool-result image parsing and tool-role multimodal detection.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 3 comments.

File	Description
lmdeploy/serve/processors/multimodal.py	Parses multimodal parts in tool messages and centralizes supported multimodal type detection.
tests/test_lmdeploy/test_content_merge.py	Adds regression coverage for tool-result image payload parsing and tool-role multimodal input detection.

---

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

[CUHKSZzxy]

This part does not apply to LMDeploy: Modality overrides __eq__ to compare against strings, so both "image" == Modality.IMAGE and Modality.IMAGE == "image" evaluate true. The parser does store modality.value, but the assertion against Modality.IMAGE is valid in this repo. I only addressed the reasonable test-surface issue by removing the heavyweight VisionModel import/collector assertion.