Show live previews during local video renders

[atomantic]

Closes #4588

Summary

• publish one atomic, job-scoped preview.png from supported LTX and MiniMax H3 MLX runner steps
• forward only the active job preview through video progress events as currentImage
• keep preview storage bounded and clean it up on build, spawn, cancel, failure, and completion paths
• cover runner geometry/projection, argument plumbing, stale-job filtering, and preview lifecycle behavior

Test plan

• cd server && npm test -- services/videoGen/local.test.js ../scripts/generate_ltx2.test.js ../scripts/generate_minimax_h3.test.js ../scripts/_runner_common.test.js
• python3 scripts/_runner_common_test.py
• python3 -m py_compile scripts/_runner_common.py scripts/generate_ltx2.py scripts/generate_minimax_h3.py
• node --check server/services/videoGen/local.js
• full cd server && npm test — 1,565 files passed, 32,819 tests passed
• cache-only local LTX smoke on the installed MLX runtime: 3 denoise steps produced a valid MP4 and one 256x256 preview.png

The full MiniMax H3 render was not run because its pinned model snapshot is not cached on this machine and the CUDA runtime is not installed; the H3 runner projection path is covered by focused tests.