Created by: zenghsh3

目前Fluid的softmax_with_cross_entropy只支持2-D的logits和label输入，但输入大于2-D的场景也是有的，建议直接支持任意大于2-D的输入。

类似于下面功能：

def softmax_with_cross_entropy_any_rank(logits, label, soft_label=False, ignore_index=-100):
    """ Expend paddle.fluid.layers.softmax_with_cross_entropy to support any rank input.
    Will do the softmax and cross entropy along the last dim of logits and label.

    Args:
        logits (Variable): The unscaled log probabilities, which is a (N + 1)-D tensor with shape
                           [d1 x d2 x ... dN x K]. K is the class number.
        label (Variable): The ground truth which is a (N + 1)-D tensor. If soft_label is set to false,
                          Label is a Tensor<int64> with shape [d1 x d2 x ... dN x 1].
                          If soft_label is set to true, Label is a Tensor<float/double> with shape
                          [d1 x d2 x ... dN x K].
        soft_label (bool): A flag to indicate whether to interpretate the given labels as soft
                           labels. By default, soft_label is set to False.
        ignore_index (int): Specifies a target value that is ignored and does not contribute to the
                            input gradient. Only valid if soft_label is set to False. Default: -100
    Returns:
        The cross entropy loss is a (N + 1)-D tensor with shape [d1 x d2 x ... dN x 1].
    """
    if len(logits.shape) > 2:
        result_shape = list(logits.shape[:-1]) + [1]

        flatten_logits = layers.reshape(logits, shape=[-1, logits.shape[-1]])
        flatten_label = layers.reshape(label, shape=[-1, label.shape[-1]])
        flatten_result = layers.softmax_with_cross_entropy(flatten_logits, flatten_label, soft_label, ignore_index)

        result = layers.reshape(flatten_result, shape=result_shape)
        return result

    else:
        return layers.softmax_with_cross_entropy(logits, label, soft_label, ignore_index)