[fx] modify the calculation of node_size in MetaInfoProp for activation checkpointing usages

[super-dainiu]

I believe this change won't affect usage of node.meta['tensor_meta'], but I have adopted the expressions in https://github.com/pytorch/pytorch/blob/f9533560ccb2d789bc78caa360d021f72f5ebc7b/torch/fx/passes/graph_manipulation.py#L82 so as to compute the memory consumptions for activation and model parameters separately. This will make it easy for us to retrieve the activation memory cost directly and correctly.

for node in graph.nodes:
    node_size = getattr(node, 'node_size')    # node_size is a NamedTuple(output_size, param_size)
    print(node_size.output_size)    # will get the memory cost of activation in bytes format
    print(node_size.param_size)     # will get the memory cost of activation in bytes format