fix(sglang): support tensor image sizes for multimodal P/D

[Chokoyo]

Summary

• Cherry-pick of fix(sglang): support tensor image sizes for multimodal P/D #11255 to release/1.3.0.
• Add the SGLang compatibility hook that derives CHW tensor image dimensions from shape[-2:] so multimodal P/D requests stay on the exact-token path.
• Preserve the no-fallback regression coverage; the release branch additionally needs the test module's explicit logging import.

Fixes DYN-3390

Original PR

• Main PR: fix(sglang): support tensor image sizes for multimodal P/D #11255
• Main merge commit: f30c9db7d9e755ef7e798ddefae23b50d3e830fb
• Backport commit: 1d63b449c9b44827fc10ac8bf3883540623115c4

Validation

• File-scoped pre-commit checks passed.
• Targeted SGLang compatibility tests passed (3 passed).
• Mypy passed for all three changed files.
• git diff --check passed.
• Qwen3-VL-2B same-GPU SGLang P/D E2E passed on the release worktree with NIXL and DYN_CHAT_PROCESSOR=sglang.
• E2E returned HTTP 200 with prompt_tokens=276; SGLang prefill ran 288 tokens.
• E2E logs contain no .height, decode+retokenize, or token drift warnings.
• Release branch CI passes.

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no bugs or issues to report.