Visibility into early Rx / late Tx

[atenart]

A few ideas here (we can split this as needed later):

• Use the NIC timestamping feature to timestamp packets in the NIC itself, so we can later compare it to the timestamp of events and report any unexpected delay. This is not that trivial because not all NIC support it, clocks need to be synced and the whole thing would modify the system config/behavior so that might not be used in all cases.
• Add a generic XDP program to report events early/late. The difficulty here is how to correlate the packet with the skb tracking logic. Part of the answer could be to hash the packet like what our OvS module is doing.
• Add a post-processing command dedicated to latency, or improve the existing ones. Also this might fit even better in a generic Python post-processing command, see Post-processing in a Python interpreter  #35.
• Improve the BPF side performances and allow to restrict what happens so we introduce as little extra latency as possible. Some options: Move the hook filtering choice to Rust #219 and Move common section handling after hooks ran #188.