Currently, Kernel.bind() inherits its 0/1 specialization from the parts of torch.compile it reuses.

This could be a problem for some use cases. We should add an option to disable this behavior.