[Feature] Native Z-Image jobs and true H3 creation progress

[yangzhuxinyzx]

Purpose

Add native Z-Image Turbo/Base image jobs and actual H3 lifecycle/denoise progress for a prompt-first creation flow. The image API supports persistent asynchronous jobs and synchronous compatibility; H3 progress uses ordered rank-zero events over the existing worker pipe, with no additional per-step device barrier or collective.

Z-Image uses local, versioned components. Measured V100 FP16 projection and AdaLN overflows are handled by retaining the required FP32 range; Base attention/residuals use FP32 with FP16 stored weights. No outlier clamping, BF16 requirement, CUDA/C++ change, GPU-policy change or Hub download fallback is introduced.

This differs from #578: that PR addresses work/efficiency metrics, while this change adds the image runtime, persistent job protocol and user-visible stage/iteration progress.

Test Plan and Result

• 53 CPU media tests passed: async/sync APIs, true counts, cancellation, restart idempotency, outlier precision, corrupt records and disk-write failure without a stalled queue.
• Native pre-commit and CI passed. Existing ABI-matching SM70 extensions from b6d91d6 were reused with PyTorch 2.10 / CUDA 12.8.
• Actual Studio frontend runs on V100 32 GB generated Turbo/Base 1024-square images and H3 text, keyframe and reference-image videos at 1344x768, 107 frames, 24 fps, with stereo audio. Playback/download/reload and media reuse passed.
• Fixed-input progress parity: Turbo/Base PNGs are byte-identical; warm H3 MP4s and raw audio are byte-identical. A first/warm audio difference up to 5.44e-7 also occurs with reporting enabled in both runs.

docs/design/native_media/CREATIVE_JOBS.md records the tested recipes and limits. Base's original 50-update recipe takes roughly ten minutes on one V100; Turbo remains the interactive default. These results do not claim exhaustive validation of every allowed output shape/duration.

AI assistance was used for implementation, debugging, tests and review. Contributions are signed off.