[fx]add gpt2 passes for pipeline performance test

[YuliangLiu0306]

This PR add some passes to walk around the bugs happened due to too flexible partition policy.
There are two main changes:

1. In new split_module pass, if a node has non-tensor data and the args of node do NOT locate in same partition as node located in, it will be moved into its args partition to avoid non-tensor data communication. To avoid circular dependency, we will also move all ancetors of this node into same partition.
2. In new split_with_split_nodes_pass, we will eliminate the unused placeholder nodes and output nodes to avoid unnecessary communication.

Only a part of changes in above pass will be added into the origin pass, and above pass will be deprecated in future.