Add: dual-slot AICPU dispatch payload and two-phase pipelining scheduler

[zhusy54]

Summary

• Introduce dual-buffer payload storage so AICPU can pre-load the next task's
  payload while AICore is still executing the current one (true pipelining)
• Add two-phase dispatch loop: issue to a pending slot while the running slot
  is occupied, eliminating idle cycles between consecutive kernel launches

Key Changes

• s_pto2_payload_per_core: extended from single-buffer to [RUNTIME_MAX_WORKER][2];
  slot selected by reg_task_id & 1u, consistent between AICPU (write) and AICore (read)
• CoreExecState: add parallel running/pending field pairs (slot_state, reg_task_id,
  subslot, dispatch_timestamp); rename executing_* → running_*
• CoreTracker: add pending_occupied_ BitStates with
  get_idle_cluster_offset_states (both slots free) and
  get_pending_only_cluster_offset_states (core running, pending slot free)
  for two-phase dispatch
• Extract decide_slot_transition() pure function to decode register events
  into SlotTransition flags; extract complete_slot_task() helper for the
  completion hot path
• AICore: add pipe_barrier(PIPE_ALL) before kernel execution and select
  exec_payload via payload + (task_id & 1u); simulation no-op fallback added
• Static assert: parity skip range keeps even parity over TASK_ID_MASK

Testing

• Existing simulation tests pass with new dual-slot scheduler
• Two-phase dispatch validated against both a2a3sim and hardware paths

[gemini-code-assist]

Code Review

This pull request introduces a dual-slot dispatch mechanism to the AICPU executor, allowing for task pipelining by tracking both running and pending tasks per core. It updates the CoreTracker and CoreExecState to manage these slots and implements a two-phase dispatch logic that prioritizes idle cores before filling pending slots. Additionally, the AICore performance collector was optimized to use a caller-maintained write index, reducing cache invalidation overhead. A fix for simulation environments was also included to prevent payload corruption. Review feedback identifies opportunities to optimize the scheduler's hot path by removing redundant bitmask refreshes.