faster by calculating R in updown and using cached R in forward

    def calc_updown(self, orig_weight):

        # this works

        R = self.R

        # R = self.R

        self.R = self.get_weight(self.multiplier())

        # this causes major deepfrying i.e. just doesn't work

        # sending R to device causes major deepfrying i.e. just doesn't work

        # R = self.R.to(orig_weight.device, dtype=orig_weight.dtype)

        if orig_weight.dim() == 4:

            weight = torch.einsum("oihw, op -> pihw", orig_weight, R)

        else:

            weight = torch.einsum("oi, op -> pi", orig_weight, R)

        # if orig_weight.dim() == 4:

        #     weight = torch.einsum("oihw, op -> pihw", orig_weight, R)

        # else:

        #     weight = torch.einsum("oi, op -> pi", orig_weight, R)

        updown = orig_weight @ R

        updown = orig_weight @ self.R

        output_shape = self.oft_blocks.shape

        ## this works