Fix: profiling buffer recycling and implicit task record emission

[ChaoZheng109]

Overhaul the profiling subsystem to fix buffer leaks, lost records, and excessive device memory allocation:

Buffer recycling (Host):

• Replace alloc/free-per-cycle with closed-loop buffer pools: completed buffers are recycled into per-type pools (recycled_perf_buffers_, recycled_phase_buffers_) instead of being freed
• Pre-allocate all PerfBuffers and PhaseBuffers at init time via PLATFORM_PROF_BUFFERS_PER_CORE and PLATFORM_PROF_BUFFERS_PER_THREAD, eliminating runtime rtMalloc calls entirely
• Reduce PLATFORM_PROF_SLOT_COUNT from 8 to 4 (sufficient with recycling)
• Process ready buffers during the expected_tasks wait phase to prevent device memory leaks when AICPU is slow to report total_tasks

Buffer recycling (AICPU):

• Replace per-cycle rtMalloc/rtFree with pre-allocated buffer arrays and free-list indices, matching the Host-side closed-loop design
• Pre-allocate phase buffers per thread and perf buffers per core at profiling init, recycling them on buffer switch

Implicit task profiling (AICPU):

• When pipelining two tasks on the same core, AICore may transition from FIN(task_A) to ACK(task_B) before AICPU reads the register. Both implicit completion paths (pending FIN and pending ACK) counted the old task but never called perf_aicpu_complete_record, losing its profiling record
• Add perf_aicpu_complete_record calls in both paths, completing the implicit task's record before the explicit task's to preserve buffer ordering

Other:

• Reduce RUNTIME_MAX_FANOUT from 512 to 128 to match actual usage
• Enable profiling setup in device_runner for both onboard and sim

[gemini-code-assist]

Code Review

This pull request implements a closed-loop buffer recycling system for performance profiling to eliminate runtime memory allocations and reduce data loss. Key changes include the reduction of maximum fanout, the introduction of recycled buffer pools, proactive replenishment of free queues, and updated profiling logic in the AICPU executor to maintain record ordering. Review feedback identified a potential stack buffer overflow during fanout copying, inefficient mutex locking in replenishment loops, overly restrictive allocation fallback conditions, and missing bounds checks for thread indices.