Investigate client WebSocket performance

[jrudolph]

In https://discuss.lightbend.com/t/low-websocket-client-performance/6907 it was reported that there might be more Streamed messages than expected which will lead to bad performance because of all the stream handling collecting full messages.

[jrudolph]

There are two things at play here (affecting both client and server parsing side):

• by default we read partial frames and stream from there on, this is probably wasteful and only useful if you use large frame sizes. It should be very easy to make configurable and change the default, maybe we should even remove the whole possibility of parsing half frames which introduces a large amount of complexity, e.g. to collect control frames.
• Preparing messages for the user is a quite complex stream (using splitAfter, prefixAndTail, concatSubstreams and various custom stages). Especially concatSubstreams is expensive. A single custom stage to encompass all of these smaller ones would help a lot, especially in the case where we always receive strict data frames (after doing the optimization in the first bullet).

[jrudolph]

3377 has some experiments for investigation but I won't continue to work on that right now.