Make HTTP/1.x server-side control flow consistent with the client-side

[idelpivnitskiy]

Motivation:

The HTTP/1.x control flow for NettyHttpServer uses a single write operation to write an indefinite stream of request/response iterations as a repeated publisher. This difference requires all our transport utilities (e.g. CloseHandler, WriteStreamSubscriber, FlushStrategy) to have special behavior to accommodate the difference.

Modifications:

• Apply repeat operator to the returned write completable;
• Deprecate SplittingFlushStrategy, it's not required anymore;
• Temporarily introduce ChangingFlushStrategy to still support changes of the FlushStrategy per request with existing FlushStrategy API;
• RequestResponseCloseHandler emits OutboundDataEndEvent on every "payload end";
• Adjust existing tests to account for the new behavior;
• Enhance some tests to use different executors for client & server, it helps with debugging;

Result:

HTTP/1.x server-side control flow is consistent with the client-side.

Note that it changes the way server-side manages demand for new writes. Instead of having a single WriteDemandEstimator for the entire connection that remembers the size of previous requests, it now has a new WriteDemandEstimator per exchange, the same way as the client. If we will observe it as a problem, we can share the WriteDemandEstimator for the entire connection in a follow up, but we should do it consistently for client & serve sides.