Card: multi-turn note updated for litert-lm 0.17.0 (past reasoning removed by the runtime; 3-turn results on the int8 file)
The multi-turn section of this card describes litert-lm 0.15.0, where past turns' <think> text stayed in the context; on 0.17.0 the runtime removes it before each turn, so the note is out of date for anyone reading the card now. This PR replaces that section with what the int8 file does on 0.17.0 in a 3-turn probe: with the default setting it answers the previous question at turn 3, with filter_channel_content_from_kv_cache off it drifts at turns 2 and 3, and the bf16 model passes with LiquidAI's own template, which keeps only the most recent turn's reasoning. The manifest's matching text is updated the same way. No model file changes; only the int8 file was re-run. If this does not fit the card, close it and I will carry it elsewhere. Thanks.
Links — the runs behind the table: https://github.com/google-ai-edge/LiteRT-LM/issues/3184 · the runtime release measured: https://github.com/google-ai-edge/litert-lm/releases/tag/v0.17.0