[WIP] streaming graph walk transition

[NSagan271]

Lets a producer partition drive a consumer's graph-walk transitions via the data stream, instead of the conductor centrally advancing them. This decouples a consumer's walk from the conductor's (lagging) view and enables PD disaggregation of the qwen3-omni Talker.

Mechanism

• PartitionDefinition.transition_source (STATE_MACHINE vs PRODUCER_TRIGGERED) declares who owns a partition's walk.
• A Connection.consumer_walk_transition callback maps streamed items to WalkTransition markers; the consumer's StreamBuffer cuts a chunk boundary at each transition so a walk never straddles two chunks.
• Per-edge produced_edge_idx/consumed_edge_idx counters round-trip through the conductor; index-based dedup (first-arrival-wins) + monotonic set_index keep the stream continuous across the prefill→decode worker seam (PD).
• qwen3-omni: Thinker streams thinker_states carrying the markers; prefill states keep the Talker in talker_prefill (KV extend), the first thinker_decode state flips it to talker_last_prefill, then the second decode state causes a transition to talker_decode. New configs/qwen3omni_talker_pd_disag.yaml splits the Talker across prefill/decode workers.
• Graph-walk transitions gated on a fully reset ("clean") worker graph.

Touches conductor/, streaming/, worker/, graph/, qwen3-omni model.

Testing

Testing in progress; tested on default qwen3 omni config, 2-GPU, thinker TP2, and a new configs/qwen3omni_talker_pd_disag.yaml splits both the Thinker and Talker across prefill/decode workers.