fix(vllm): expand Kimi K3 media pads at the worker

[furionw]

Cherry-pick of #12394 onto release/1.4.0-kimi-k3-dev.1.

Original PR: #12394
Merge commit: 0d5034a

Test Plan

• Same as fix(vllm): expand Kimi K3 media pads at the worker #12394; no additional changes in this cherry-pick.

[datadog-official]

🎯 Code Coverage (details)
• Patch Coverage: 100.00%
• Overall Coverage: 29.29% (-6.60%)

_{This comment will be updated automatically if new data arrives.
🔗 Commit SHA: fe40910 | Docs | Datadog PR Page | Give us feedback!}

[devin-ai-integration]

Devin Review found 2 potential issues.

[furionw]

The failures themselves are unrelated:

• mypy: error in untouched components/src/dynamo/sglang/register.py.
• rust-gpu: compile errors in untouched lib/llm/.../loader.rs.
• vLLM amd64/arm64: dependency-contract tests detected new ec_transfer_params and locality fields; the PR’s added request-processor tests did not fail.
• vLLM multi-GPU: unrelated Qwen text-serving timeout/server-start failure.
• deploy-operator: conversion webhook connection refused.
• lychee: existing broken documentation links.
• The three status checks above are merely aggregate failures from those child jobs.