Fix: skip local ready buffer on PENDING dispatch to reduce same-thread re-claim

[zhusy54]

Summary

• Add skip_local parameter to PTO2SchedulerState::get_ready_tasks_batch() and AicpuExecutor::pop_ready_tasks_batch()
• When a thread dispatches PENDING tasks (is_pending == true), it now passes skip_local=true to avoid immediately re-consuming tasks it just released into the local ready buffer
• This gives newly-released downstream tasks a window to be flushed to the global queue and picked up by other threads, preventing greedy same-thread re-claim

Testing

• All tensormap_and_ringbuffer L3 hardware tests passed (test_l3_child_memory, test_l3_group, test_l3_dependency)
• Pre-commit hooks passed (check-headers, clang-format, clang-tidy, cpplint)

[gemini-code-assist]

Code Review

This pull request introduces a skip_local parameter to the task scheduler to improve task distribution across threads by allowing a window for local tasks to be flushed. It also refactors the pop_ready_tasks_batch function to accept profiling counters as references. A critical issue was identified in the profiling logic where the pop_hit counter is incremented by one per batch instead of the actual number of tasks popped, which will result in incorrect performance metrics.