Enhance sinks to make them batchable

[hunterlxt]

close #415

Optimize send_all using sink in grpc-rs. By buffering the message to be sent to determine whether there is still message in the stream, and add the buffer_hint flag to indicate follow-up messages.

After a simple test, this does improve the QPS by 50% when you use send_all method.

Before
======== test_client_stream_send_split
QPS: 13626.411713542033
======== test_client_stream_send_all
QPS: 13458.8795676736
======== test_server_stream_send_all
QPS: 21513.004479575207

After
======== test_client_stream_send_split
QPS: 13632.78241828707
======== test_client_stream_send_all
QPS: 20634.743684214573
======== test_server_stream_send_all
QPS: 21282.527286920154

Tcpdump capture file size (only send_all test)
Before：22M
After：2.6M

Related test code: https://github.com/hunterlxt/grpc-rs/blob/xt/bench-sink-over-master/tests-and-examples/examples/hello_world/client.rs

Additional benchmark with standard grpc driver (before vs after)

rust_protobuf_async_streaming_from_client_100msg_unconstrained
QPS: 12058.0
QPS: 41914.1
rust_protobuf_async_streaming_ping_pong
QPS: 18315.0
QPS: 20409.9
rust_protobuf_async_streaming_qps_unconstrained
QPS: 589875.8
QPS: 584359.6

[BusyJay]

Cool, how does grpcio benchmark say?

[hunterlxt]

Cool, how does grpcio benchmark say?

The updated results are in the description. BTW, I read the benchmark code, the client does not use the send_all method.

[BusyJay]

If adapted to send_all, how does it performs?

Since now std::future is used, async/await is the preferred style. It will be less useful if the optimization can't work with that.

[hunterlxt]

If adapted to send_all, how does it performs?

Since now std::future is used, async/await is the preferred style. It will be less useful if the optimization can't work with that.

send_all is a method of SinkExt. futures::sink and std::future do not conflict with each other. For users, the method of std::future is not directly used, but the method in SinkExt of futures::sink.

Above is the background. The semantics of send() is to send and ensure that this message is indeed successful. But if the user really wants to send a stream, he should choose the send_all method, and our sink has not been optimized for this.

[BusyJay]

Can you give an example about how to utilize the optimization with async/await?

[hunterlxt]

The previous benchmark in grpc-rs did not implement the stream_from_client bench. I added it and performed the bench. The result has been updated. cc @BusyJay

[hicqu]

rest LGTM.

[BusyJay]

Rest LGTM

[hunterlxt]

@BusyJay @hicqu PTAL