Misc. Benchmark Updates

[IanWold]

• PAYLOAD_BYTES wasn't being used for anything, now that does vary the bytes/payload
• The server semaphore was being recreated on each publish, now it's a single instance
• Batches are dispatched before the last ones finish receiving; CONCURRENCY controls this
• Add flag to be able to control payload strategy without needing to modify code
• Adds variable to control number of repeats
• Add "shared load" server to simulate scenarios where the backplane is not dedicated as a backplane; either a Redis that is also a cache or a PG that is also a regular database
• Allow configuring anywhere from 2-10 servers, and configure number of clients/server rather than a strict number of clients.
• Update concurrency and between-run draining logic
• LLM-generated shell script to run benchmarks on cloud provider (railway) instead of locally