complete distributed lookup table

[jacquesqiao]

Tasks

• distribute lookup table support async training Fix and optimize async distribute lookup table #10973
• add auto_grown_mutex for selected_rows add auto_grown_mutex for selected rows #11012
• run dist lookup table on dlnel cluster
• run dist lookup table on paddle cloud
• support use lookup_sparse_table at local training to support OOV(out of vocabulary) id.
• suport variant method to initialize non exists ID in the table(Guass, Uniform or zero)

support deep and wide model

• concat_op support data layer as input concat support data as input #11527
• support Ftrl optimizer add ftrl optimizer #11547
• refine prefetch to support multi thread prefetch Refine prefetch #11366
• add a merge_ids op to replace concat op to merge the result of distribute lookup table. Add merge_ids_op #11342
• support auc calculate fix auc #11574
• support save/load dist look up table on pserver side. @seiriosPlus Checkpoint M2: lookup table checkpoint #11490
• fix adam op for selected rows fix adam op for selected rows #11765

[Yancey0623]

we can also support OOV(out of vocabulary) with local training, and initialize the parameter of the unknown ID as 0, Gauss or uniform random.
Related issue #11361

[jacquesqiao]

@Yancey1989 Yes, we can support to use lookup_sparse_table at local training. Add a task in the task list.