Watermark based backpressure for webserver

[danielkec]

Investigate use of backpressure for not immediately flushed datachunks as suggested by @olotenko :

Count the bytes that go through here, and allow up to a given amount to be buffered. Eg onNext counts how many bytes were in the buffers it has seen, and requests one more chunk if it has seen below watermark. Then channel.write completion Future subtracts the bytes it has just written. If this results in crossing the "watermark", then request 1. (because onNext stopped requesting more)

Watermarking approach is more predictable. Just deferring to channel.write completion would work poorly, if in the common case socket write completion does not happen during onNext, but will work better than watermarking, if there is this guarantee. (Ie that onNext attempts to call channel.write, through a few calls of intermediate routines, and that Netty completes the future before onNext returns, if the write fully fits in socket send buffers)

#3108 (review)

PR's

• 4x Watermarked response backpressure 4x #5063
• 3x Watermarked response backpressure 3x #4724
• 2x Watermarked response backpressure 2x #5062