Windows CUDA ops and 5090 low-VRAM load path

[JonEcho]

Scope

Windows load path for JoyAI-Video-Edit on an RTX 5090 (32 GB).

• Build joyomni_ops with MSVC and CUDA 12.8 (deploy/build_joyomni_ops.ps1, deploy/install_vs2022_cpp.cmd, deploy/requirements-windows.txt).
• Skip Windows VAE inductor compile unless JOYOMNI_VAE_COMPILE=1.
• Cap the CUDA allocator with JOYOMNI_VRAM_CAP_GB.
• Fall back when CUTLASS FP8 GEMM is missing or broken so 720p DiT still loads.
• Use asyncio instead of uvloop on win32.
• Raise holder idle timeout to 1800s so a long restyle clip does not drop the websocket.

Stack

This is PR 1 of 2.

Child: JonEcho#1 (file-restyle-loop-splice; base is this branch).

Verification

• git diff --stat origin/main...HEAD shows 13 files, +241 / -65.
• Local code_rules_gate --staged reported 25 introduced hits on this fork. Those hits match existing JoyAI style (imports inside functions, print(), large create_app). This contribution does not rewrite that style.
• This session did not re-run GPU serve. A prior 720p restyle on this machine used JOYOMNI_VRAM_CAP_GB=26 and VAE on CPU.

Risks

• FP8 fallback is slower than a working CUTLASS kernel.
• Idle timeout 1800s applies on every platform, not only Windows.
• build_joyomni_ops.ps1 still assumes VS 2022 Community vcvars64.bat and a repo-root .venv.

Follow-up

Merge stays with the maintainer. This pull request is draft.