Mux: improve egress fairness

[coot]

We need to improve the fairness of the egress side of the multiplexer. See #5261, why this is needed.

[coot]

We realised that the observed unfairness was a misinterpretation of the results. TLDR: Mux, as it is currently implemented, is fair when bandwidth is limited and efficient when an excess of bandwidth is available.

The following graphs were produced using #5293 simulation, in which we have two mini-protocols. Both make an initial request, after which the server continuously pushes data payloads over the multiplexed TCP connection. The two mini-protocols only differ by the amount of data which is sent in each response.

Protocol Type	Protocol Number	Block Size
Praos	0x8002	10xSDU
Leios	0x8003	1000xSDU

SDU=12kB, bandwidth = 10mbit/s

But when an excess of bandwidth is available, the egress side for the protocol which sends larger payloads becomes more efficient:

Protocol Type	Protocol Number	Block Size
Praos	0x8002	10xSDU
Leios	0x8003	1000xSDU

SDU=12kB, bandwidth=500mbit/s.

The efficiency stems from the fact that Wantons in mux egress have only a soft limit, whereas there's no hard limit; for example, a mini-protocol can place an arbitrary amount of data above its soft limit into its Wanton if the Wanton currently has less than the soft limit. For protocols which smaller chunks of data, as for the Praos protocol, it means that the egress side will be more often blocked, getting bytes from the application layer (mini-protocol) and adding them to its Wanton, while the Leios mini-protocol can make progress sending data (cutting full SDUs from its Wanton and sending them over the network).

If we introduce a hard limit, the efficiency of Leios protocol is capped:

Protocol Type	Protocol Number	Block Size
Praos	0x8002	10xSDU
Leios	0x8003	1000xSDU

SDU=12kB, bandwidth=500mbit/s, MUX with soft & hard limits.

Our current understanding is that there's no compelling reason to introduce a hard limit, as it doesn't improve the throughput of the Praos protocol in this experiment.