Merge selectedrows dynamically on the server site

[Yancey1989]

#7838 implement batch barrier feature, and we also need to implement merging SelectedRows in server site dynamically, just because:

1. After the distributed transpiler, a SelectedRows variable block maybe contains an empty Tensor, and we don't need to send the empty Tensor in SendOp.
2. For the server site, we can not use an Op to merge the received SelectedRows, because the Input size should be a definite value, but the number of received SelectedRows variable is different between each batch.

So I think we also need to implement the following feature:

1. Send Op don't need to send the variable which was not initialized.
2. Merge SelectedRows in server site.

[typhoonzero]

1. Do not generate "empty variables" in transpiler. Since we have barrier we can just send any count of variables and signal the server the send is complete.

[Yancey1989]

FROM @typhoonzero

Do not generate "empty variables" in transpiler

Yes, but I think it's not exactly, for the sparse update, the transpiler would split a SelectedRow into multiple SelectedRows by the attribution height_sections, for each batch, SelectedRows.value() could be empty.

So I think that SendOp doesn't need to send an empty variable is Okay.