KV Cache Backed Branches

Introduces persistent conversational branches using transformer inference state, allowing users to explore multiple continuations while preserving prior computation. This improves AI productivity and interactive workflows.