Introducing gpt-oss

Publishes the architecture table (layers, total/active params, total/active experts, context) and states alternating dense and locally banded sparse attention, GQA group size 8, RoPE, 128k native context.