Add: require_sync_start for atomic SPMD block launch

[poursoul]

Introduce a sync_start mechanism that forces all blocks of an SPMD task to be dispatched atomically before any can begin execution.

Submission layer (pto_submit_types.h, pto_orchestrator.cpp/h):

• Add LaunchSpec::require_sync_start and active_mask bit-3 flag
• Add pto2_core_mask() / pto2_requires_sync_start() helpers
• Validate block_num < total resources at submit time to prevent deadlock
• Fix total_required_subtasks to use pto2_core_mask (strip flag bits)

Scheduler drain protocol (aicpu_executor.cpp):

• Three-phase drain: ack barrier → global resource check → exclusive dispatch
• Elected thread verifies global idle resources before dispatching; if insufficient, all threads return to completion polling and retry
• Non-elected threads spin-wait during dispatch, giving the elected thread exclusive CoreTracker access (no data race on core_states_)
• Track active_sched_threads_ separately from thread_num_ so orchestrator threads that have not transitioned to scheduling do not block the ack barrier

SPMD dispatch refactor:

• Extract dispatch_block_to_cluster / dispatch_mix_block_to_cluster
• AIV path uses count_idle_aiv_cores for accurate resource counting

Test examples: spmd_sync_start, spmd_sync_start_aiv, spmd_sync_start_edge, spmd_sync_start_stress, spmd_starvation

[gemini-code-assist]

Code Review

This pull request implements the sync_start mechanism for SPMD tasks, ensuring atomic launching of all blocks. It introduces a drain protocol in the AicpuExecutor to coordinate resource allocation across scheduler threads and prevent starvation. The submission logic was updated to handle the require_sync_start flag, and a suite of golden tests was added to verify the feature's correctness and robustness. Feedback identifies a potential undefined behavior in bitwise shifts for thread masks and suggests adjusting the deadlock check to allow tasks that require exactly the total available resources.

[ChaoWao]

#441