[DeepLiftShap] Callable background data

[ghkim9213]

          What do you think of also supporting `callable` methods as a parameter? In such case, we would be able to provide a default random generator based on input with something like:

class DeepLiftShap(Explainer):
    SUPPORTED_MODULES = [Convolution]
    
    def __init__(
        self,
        model: Module,
        background_data: Optional[Tensor, Callable[[Tensor, Tensor], Tensor]] = None,
    ):
        if background_data is None:
            background_data = lambda x, y: torch.rand_like(x)
        if torch.is_tensor(background_data):
            background_data = lambda x, y: background_data

        self.background_data = background_data

Originally posted by @enver1323 in #151 (comment)