[llama4] fall back to for-loop based MoE if not on SM90 or later

[tianyu-l]

as titled

also fixing a breakage caused by #1086

[lessw2020]

I feel like this could be made more general, otherwise you are going to have an evergrowing fleet of specific functions and users have to check if said specific version exists...
Instead how about a simple api with major minor verification?

is_sm_version_or_higher(major, minor=0)

and thus checking SM90 for example is just

is_sm_version_or_later(9, 0)

this means adding things like Blackwell (SM100, SM120) etc. do not need a new function - they can all use the same thing in the future rather than writing specific version checks.

rough code...did not test but to give an idea:

def get_cuda_capability():
    if torch.cuda.is_available():
        return torch.cuda.get_device_capability()
    return None

def is_sm_version_or_higher(major: int, minor: int =0)-> bool:
    """
    Returns:
        bool: True if GPU meets or exceeds the specified version
    """
    capability = get_cuda_capability()
    return capability is not None and capability >= (major, minor)

[lessw2020]

looks good.
Left a comment inline that I do think switching from a series of specific version check functions is more elegantly expressed as a single api that takes in (Major,Minor) and avoids future technical debt since this would scale to future versions intrinsically.
But can be a future PR as well, will work fine as is.