diffusion-ci: restore main-compatible Sana H100 GT

[BBuf]

Summary

• Restore the three H100 sana_wm_ti2v_1gpu consistency frames to the blobs used before 88b17d4.
• Keep the Qwen-Image fused-QKV H100 ground truth added by diffusion-ci: refresh Qwen-Image H100 GT for fused QKV #5 unchanged.

Why this is needed

88b17d4 (Update SANA TI2V H100 consistency ground truth) captured output from the unmerged warmup-calibrated auto-residency experiment in sgl-project/sglang#35335 / the stacked sgl-project/sglang#36703 work, rather than output from SGLang main.

That experiment introduced post-warmup residency changes in 7f66f1c and enabled them in diffusion CI in dfb2cfe by removing SGLANG_DIFFUSION_DISABLE_AUTO_RESIDENCY=1. On the Sana H100 case, the runtime changed placement after warmup:

transformer_2: component-offload -> resident
transformer: component-offload -> resident
text_encoder_2: layerwise-offload -> resident
text_encoder: resident -> component-offload

The two-stage Sana path was not numerically invariant under those transitions. The old-GT comparison in run 33124193263, partition 3 reported:

min CLIP similarity: 0.8663
min SSIM:           0.6270
min PSNR:           19.3263
max mean abs diff:  13.0355

Those are exactly the metrics observed in the opposite direction when the normal mainline Sana output is compared with the 88b17d4 images in sgl-project/sglang#36680. The experimental PR subsequently documented the problem and opted Sana out with supports_auto_residency = False in 0834013; it later moved its CI-data pin back from 88b17d4 to its parent in 137286c.

Because sgl-project/sglang#35335 is still unmerged, SGLang main never adopted the path that produced the replacement images. Current mainline Sana runs remain bit-exact with the pre-88b17d4 ground truth. The experimental images nevertheless remained in the shared ci-data history and were unintentionally pulled into sgl-project/sglang#36680 when that PR advanced the repository-wide revision to include the unrelated Qwen-Image GT from #5.

This PR removes that cross-PR contamination instead of weakening Sana thresholds or blessing output from an unmerged, already-disabled numerical path.

Validation

• Only the three H100 Sana frames change.
• Their Git blob IDs exactly match f8c4481d1fa2fa1b9efdfe6831d9c60dbf386b4a, the parent of 88b17d4:
  • frame 0: e47c0559f4f9545a92be9f1417208549db3ecd04
  • frame mid: 68706eb5ec3f03b439e4f23df7f89f91d4a1ddce
  • frame last: b2f4938b90be6731b4ba5e9400267137274d3e48
• The Qwen-Image fused-QKV GT from diffusion-ci: refresh Qwen-Image H100 GT for fused QKV #5 remains unchanged.

Follow-up

After this merges, sgl-project/sglang#36680 should update SGL_TEST_FILES_CI_DATA_REVISION to this PR's merge commit and rerun the multimodal-generation CI.

References:

• [diffusion] Auto residency: per-pipeline containment fixes and CI baselines (4/4) sglang#35335
• [diffusion] Plan reversible residency before warmup sglang#36703
• [Diffusion] Optimize Qwen-Image TP collectives and attention sglang#36680
• ci-data commit 88b17d4fca324f9994d6f9f58495ed3882700734