Every byte counts: ARCv3 and the case for cross-region RL

Introduced ARCv3, a new compressor for BF16 weight updates in Reinforcement Learning (RL) training. ARCv3 reduces delta payload size by nearly 50% compared to ARCv2 (0.19% of original BF16 weights vs. 0.36%), achieving lossless reconstruction. It exploits the asymmetry in weight updates, prioritizing mantissa-only changes and packing rarer exponent/sign changes. ARCv3 also uses parallel general-purpose compression algorithms for further optimization. The system is integrated into the Fireworks Training API via `fireworks-delta-compression`.