ci: add post-merge K3 two-host smoke for feat/k3_dev

[lhcezx]

What changed

• add a push workflow for every update to feat/k3_dev
• enqueue the dedicated Aone K3 two-host smoke without GitHub-side cancellation or concurrency replacement
• pass repository, branch, commit SHA, GitHub run metadata, and a stable idempotency request ID
• add one-shot Aone status parsing with distinct accepted/final states and task links

Why

feat/k3_dev currently accepts merges without an automated full-model Kimi K3 regression run. The heavy TP8/EP8 PD smoke requires trusted two-host scheduling on fixed GPU workers, so GitHub only submits the task while Aone owns FIFO resource waiting and result reporting.

Runtime topology

• Prefill: 106
• Decode: 142
• smoke: example/k3/kimi_k3_full_model_two_host_pd_smoke.sh
• suite: SMOKE_SUITE=all
• load method: fastsafetensors
• GitHub check: K3 CI / two-host smoke

Validation

• bash -n
• ShellCheck 0.11
• Actionlint 1.7.12
• git diff --check
• Aone YAML syntax validation
• trusted controller unit tests: 8 passed
• 106/142 read-only image, checkpoint, RDMA, lease, task-container and cleanup validation

Rollout prerequisites

This PR intentionally remains draft until the dedicated Aone pipeline and narrow secrets are provisioned and a manual full smoke passes while both GPU hosts are idle.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.

---

罗昊成 seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}