Refactor aicpu logical

[poursoul]

• Replace std::mutex with spinlocks and bitmask circular queues in
  AICPU scheduler ready queues; remove redundant atomic counters
• Add orchestrator ready queue (SPSC ring) to push early-return ready
  tasks directly, replacing O(N) readiness scan in scheduler
• Stack-allocate tensormap lookup results (PTO2LookupResult) to avoid
  per-lookup heap allocation from std::vector
• Inline Tensor copy ctor/operator= into header; skip memset on task
  ring slot allocation; bulk memcpy params in submit_task
• Add PTO2_SPIN_PAUSE_LIGHT() (yield without sched_yield) for
  tight spinloops on aarch64/x86_64
• Add DEV_ALWAYS log level for diagnostics; change diagnostic reports
  from DEV_ERROR to DEV_ALWAYS, per-task logs from DEV_INFO to DEV_DEBUG
• Add profiling instrumentation to scheduler, orchestrator, and host
  runtime init
• Vectorize paged attention golden computation across batch dimension
• 修复64batch卡死的问题，暂时将window大小改成65536，后面再支持task ring的环形分配
• replace pa golden