test(serve): cover user images across three HTTP APIs

[silencelamb]

Motivation

Equivalent user-image requests should preserve image contents, text order, and conversation history across /v1/chat/completions, /v1/messages, and /v1/responses. Use a small shared set of scenarios to check these invariants across all three entry points.

Modification

Add tests/test_lmdeploy/serve/test_user_images.py with three representative PNG scenarios:

Scenario	Image source	Shared assertions
Image-only user message	Data URL	Image reaches preprocessing without requiring text
Multiple images interleaved with text	HTTP URLs	Decoded image contents and text/image order are preserved
Images in separate conversation turns	Data URLs	Roles, assistant text, and image history are preserved

Each scenario runs through all three APIs in both non-streaming and SSE modes, also checking successful completion and session cleanup: 18 HTTP cases.

Six additional checks use protocol conversion and shared preprocessing directly to ensure corrupt image data is rejected and text-only models reject images, once per protocol.

The suite uses production routing, protocol conversion, media loading, and image decoding. Generation and remote HTTP transport are simulated; no model weights, GPU service, or external network access are required.

Test scope

This file owns representative cross-endpoint invariants, not a Cartesian product of image formats, sources, and conversation shapes. PNG/JPEG/WEBP decoder coverage, SDK compatibility, image parameters, and detailed protocol validation remain in the existing endpoint suites, including openai/responses/test_images.py.

A few successful image flows intentionally overlap with the focused Responses regression tests so that #5043 remains independently validated. Tool-result images and their three-endpoint regressions belong to #5044.

BC-breaking (Optional)

None. The incremental change adds one test file without changing production code or dependencies.

Validation

• Complete serving suite at 70689117: 510 passed (python -m pytest tests/test_lmdeploy/serve -q), including all 24 shared image checks.
• Applicable pre-commit hooks and git diff --check passed.
• No GPU service was started. Earlier real GLM-5.3-Flash validation is recorded in [Fix] Handle user image inputs in Responses API #5043; these tests do not validate model generation.

Dependency and review scope

This PR depends on #5043, which enables ordinary user images in Responses. It does not depend on #5044.

Until #5043 merges, the upstream main diff also includes that prerequisite. The incremental test-only comparison adds only tests/test_lmdeploy/serve/test_user_images.py. This branch should be rebased onto upstream after #5043 merges.

Checklist

• Applicable pre-commit hooks and whitespace checks passed.
• The complete serving test suite passed.
• No new dependency or downstream minimum-version requirement.
• Test scope and simulated components are documented in the test module.