ci: qualify immediate two-GPU execution on E1 PRs

[richardoo-707]

This draft qualifies immediate PR-triggered execution on the gpu-h20-5 self-hosted runner. Opening or updating the internal E1 branch PR selects two idle GPUs, runs a pinned-container NCCL/DDP smoke test, and records the tested merge SHA, selected devices, and numerical result in Actions.

The job fetches the official GitHub source archive at the immutable PR merge commit and extracts only its four E1 scripts into a fresh temporary directory. This accommodates the node's unreachable github.com Git endpoint while codeload.github.com remains reachable. The archive SHA256 is recorded; downloads and test execution have explicit time limits.

The initial trigger is restricted to this branch and author. It uses manual resource authorization, does not contact the reservation website, fails when two idle devices are unavailable, and cleans up its own test container. This is infrastructure qualification, not ColossalAI version-compatibility coverage.

Validation:

• Six event/resource-selection/result-validation unit tests passed locally.
• Workflow YAML, Python formatting, and whitespace checks passed.
• Real two-GPU immediate and scheduled smoke tests passed before this PR.
• PR update automatically triggered run 35446857562, which passed all steps on colossalai-e1-h20-5. GPUs 3 and 4 passed matrix multiplication, NCCL all-reduce, DDP gradients, and SGD/equal-weight checks. The container was removed and both devices returned to 4 MiB/0% utilization.
• Tested head: 25a8d2c; tested merge: e9037db. GPU execution: 2026-09-19 13:51:12–13:51:20 UTC.

Keep this qualification PR unmerged pending review of its Actions results and resource-use policy.