dmesg was showing this messages:
[Thu Aug  7 14:05:26 2025] net_ratelimit: 4328 callbacks suppressed
[Thu Aug  7 14:05:26 2025] neighbour: arp_cache: neighbor table overflow!
[Thu Aug  7 14:05:26 2025] neighbour: arp_cache: neighbor table overflow!
[Thu Aug  7 14:05:26 2025] neighbour: arp_cache: neighbor table overflow!
[Thu Aug  7 14:05:26 2025] neighbour: arp_cache: neighbor table overflow!

and the machines were becoming inaccessible. increase the arp cache size
fixes this.

we were hitting conntrack limits when opening lots of connections and
sending UDP packets to many different hosts. this resulted in TCP
packets getting dropped which would manifest itself as errors when
connecting or timeouts and when sending UDP packets using `sendto` it
would fail with permission denied error. disabling conntrack fixes all
of these problems.

there were messages similar to:
HTB: quantum of class 10020 is small. Consider r2q change.
that showed up when brining up the network. this commit fixes that.

the env var OAR_P2P_CONCURRENCY_LIMIT limits the number of parallel
"operations" being done on the cluster machines. so, if it is set to 3,
then we only work on 3 machines at time. setting to 0 means unlimited.

currently the shell script used to list 10.0.0.0/8 range of addresses on
a machine would fail with exit code 1 if no addresses were present in
that range (i.e. grep did not match anything). this fix just makes sure
that command always returns exit code 0.