DiffusionGemma Encode (KV Cache)

Mints or advances a DiffusionGemma KV-cache from raw-encoded text (context, not a chat-templated prompt/turn). The optional kv_cache input is a dual door: leave it UNWIRED to mint a fresh cache from text alone (IN-1); WIRE an existing kv_cache in to ADVANCE it with the newly-committed text (IN-3). Same node, one call — the presence of a wired cache is the only thing that switches mint vs. advance. Feed the output cache into DGemmaDenoise's kv_cache input to condition a run on it — the turn text itself belongs on DGemmaDenoise's prompt widget. Caution: if DGemmaDenoise's decoder composes a prompt onto this node's output cache (ADR-CDG-024), the cache object is grown in place; ComfyUI's node-result caching can then silently hand a later run the already-grown cache. If you see the wrong/stale turn echoed back, invalidate this node (change its input, or force re-execution) rather than reusing its cached output (#265).