Introducing the Atlas Inference Server and Engine

Source is released. @tbraun96 and team showed their work. Credit where credit is due.

All of the Dockerfiles are based on Nvidia CUDA images, including the multi-target, you can compile and build it completely locally.

Juciest target looks to be Qwen3.5-35B-A3B; would be nice to have MTP in the Qwen3.6 version.