Accept launch parameters as arguments to forward kernel

[horizon-blue]

So that we can tune them on the fly...

Usage:

out_img = render_fmbs(
    scene,
    blender=blender,
    confidence=confidence_model,
    intr=intr,
    extr=extr,
    img=image,
    grid_size=dim3(16, 16),
    block_size=dim3(16, 16),
)

e.g. For 256x256 image on 5090 (the timing isn't really meaningful here other than showing that this works)

[horizon-blue]

I'm adding the -Xptxas -O3 flag because in the past it's been giving me better performance... If you're running into build issue, you might need to clear out the build cache once (or also feel free to delete this)