Deterministic execution and rack power management

Nvidia's Groq-derived LPX rack posts 3,431 tokens/sec on 128GB of SRAM

The first third-party benchmark of the LP30 rack came in at roughly four times the next-fastest public endpoint, measured one request at a time on a model small enough to fit.