Simulations to improve gossipsub performances & reliability

[Menduist]

Use testground to:

1. Fix bugs in our gossipsub implementation [in progress]
2. Experiment with scoring & attacks [in progress]
3. Compare it to other implementations, improve it if required

I'll use this issue to track my progress on the matter.
Here is my working branch: vacp2p/libp2p-test-plans@46e2269...status-im:libp2p-test-plans:pubsub2

To be able to use wireshark, I've started to implement #821

I've quickly noticed that Naggle was causing issues, opened #822 which is waiting on chronos to be merged.

Also found a bug in yamux #823

And obviously a few bugs in gossipsub #827

I can simulate ~200 nodes on my laptop, or rent beefy AWS servers for more when required.

This is an example of a run with 100 nodes, 30 publishers, 22 connections per nodes, and a fixed 100ms of RTT.
After 20 seconds, a massive blackhole attack start, where the 70 remaining nodes are swallowing every messages silently.

With a pretty simple scoring setup (only relying on firstMessageDeliveries & opportunisticGraft), the effect on latency is quite minimal, and we don't get any message losses, which is very good. I'm hoping to generalize to scoring setup, and test it in a wide variety of scenarios to make sure it holds.

[Menduist]

Today, I did a simulation to understand the impact of the amount of "outbound only" nodes in a network.
This is critical to waku, since it's going to be a network with a majority of nodes "in the wild" (compared to "in a DC"), with a good chunk of nodes relying on hole-punching, that can sometimes fail.

Here are some results:

I'm a bit surprised to already see an effect above 70%, something to investigate to figure if it's a bug or an emerging behavior