Add early CUDA availability check to mjwarp-viewer

[ritikkumar1212]

mjwarp-viewer currently prints "CPU-only mode" and later crashes with
RuntimeError: Must be a CUDA device.

This PR adds an early CUDA availability check after wp.init(), so the viewer
fails fast with a clear, actionable error message on machines without NVIDIA GPUs.

This improves UX for CPU-only users and avoids a deep Warp stacktrace.

[hartikainen]

I think it would be good to support non-CUDA devices. warp-lang README says:

Warp can run on x86-64 and ARMv8 CPUs on Windows, Linux, and macOS.

I'm pretty close to getting this to work with the MJX frontend in google-deepmind/mujoco#2948. I'm pretty sure that warp would work pretty much out-of-the box.

[ritikkumar1212]

Thanks for the feedback!

My understanding was that wp.ScopedCapture() currently requires CUDA, which is why I added the early guard to avoid the late runtime crash.
I’m happy to adjust this to allow CPU execution if we can bypass graph capture on non-CUDA devices (or fall back to the C engine / non-captured Warp path).
Could you suggest the preferred behavior for CPU? For example:
skip ScopedCapture and call mjw.step directly?
or auto-switch to EngineOptions.C?

Happy to update the PR accordingly.

[hartikainen]

I'm no expert on this, but the warp-lang=1.11.0 Scoped Devices documentation at least provides an example for ScopedDevice("cpu"):

with wp.ScopedDevice("cpu"):
    # alloc and launch on "cpu"
    a = wp.zeros(n)
    wp.launch(kernel, dim=a.size, inputs=[a])

[thowell]

@ritikkumar1212 thank you for contributing to mujoco warp!

#1107 should enable the viewer to be compatible with device=cpu

[thowell]

with #1107 merged the viewer should now work on cpu with --device=cpu

[ritikkumar1212]

That's great