PDU loop improvements and cleanup

[jamwaffles]

• Frame wakers can't be accessed concurrently, so the spin::RwLock around them is removed. scratch that - MIRI is unhappy. I still believe they can't be accessed concurrently, but at least we're protected now. Performance didn't really change either so that's nice.
• Replaced cookie_factory with my own small set of functions. These can panic instead of returning Result but are guarded by the ethernet frame creation code checking buffer sizes.
• Massively improve unrealistic throughput benchmark (which doesn't use a network interface so take with a pinch of salt) by changing to a simpler block_on method: theoretical max perf up to 15M packets/sec or 120 MiB/sec on my Windows 11 5950X machine. Throughput with a real network interface will be a lot lower I imagine but it's nice to see we can get over the 100Mbit 100Base-T limit in contrived cases.