feat: add warm sequential video render batches

[atomantic]

Summary

Add a 1–20 render batch control for local MiniMax H3 MLX models. One Python process retains the pipeline, model weights, adapters and prompt embeddings while rendering separate videos sequentially. Blank/dice seeds draw independently for every output; an explicit seed increments for each render, including zero.

Keep the batch in one cancelable queue slot, record every finished video with its actual seed, and preserve/show completed outputs after a later failure or cancellation. Reject overflowing seeds, unsupported runtimes, remote/cloud targets, chained clips and scene-delivery batches before dispatch. Other runtimes display an explicit availability explanation.

Test plan

• Passed video service/route suites, queue projection and generated-manifest checks, and Python runner contract tests.
• Passed relevant client form, submission, gallery-refresh and compose-while-busy tests.
• Verified partial failure, cancellation, incomplete output and first-output watchdog behavior using a controlled render process.
• Passed lint for changed client files and the client production build.
• GPU execution and performance were not benchmarked on the live instance.