upgrade async distributed training in pscore

[zhaocaibei123]

PR types

Function optimization

PR changes

Others

Describe

upgrade async distributed training in pscore：

1. fuse lookup_table to distributed_lookup_table, fuse lookup_table_grad to distributed_push_sparse in trainer_pass
2. push sparse grad by distributed_push_sparse op using fleet->comm->brpc_client
3. push dense grad by send op using fleet->comm->brpc_client(send op behavior changed, now we use send op only to push dense grad)
4. upgrade lookup table from CommonSparseTable to MemorySparseTable in server to support incremental storage for embedding
5. add ShowClickEntry to push show/click to server, which is used for MemorySparseTable.

We now support large scale sparse model inference online used widely in recommendation/search/ad by large scale kv service, which loads incremental sparse model. If you want to use this, set entry and embedding like this, then you can dump incremental sparse model.

        import paddle
        paddle.enable_static()
        sparse_feature_dim = 1024
        embedding_size = 64
        shows = paddle.static.data(name='show', shape=[1], dtype='int64')
        clicks = paddle.static.data(name='clk', shape=[1], dtype='int64')
        input = paddle.static.data(name='ins', shape=[1], dtype='int64')
        # show/clk is var_name of show/click data
        entry = paddle.distributed.ShowClickEntry("show", "clk")
        emb = paddle.static.nn.sparse_embedding(
            input=input,
            size=[sparse_feature_dim, embedding_size],
            is_test=False,
            entry=entry,
            param_attr=paddle.ParamAttr(name="SparseFeatFactors",
                                       initializer=paddle.nn.initializer.Uniform()))            

[paddle-bot-old]

✅ This PR's description meets the template requirements!
Please wait for other CI results.

[paddle-bot-old]

Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

[yaoxuefeng6]

LGTM

[chenwhql]

LGTM for _init.py