Refactor: unify PMU/L2Perf/TensorDump collectors on shared profiling framework

[ChaoZheng109]

Summary

• Introduce src/a2a3/platform/include/host/profiling_common/ with ProfilerBase<Derived, Module> (CRTP-based mgmt + collector thread orchestration) and BufferPoolManager (pre-registered device buffer pool, dev↔host pointer mapping).
• Rewrite PmuCollector, L2PerfCollector, and TensorDumpCollector on top of the shared framework, collapsing three near-identical control flows into one and dropping ~2000 lines of duplicated .cpp code.
• Reorganize profiling docs: move pmu-profiling.md from src/{a2a3,a5}/docs/ to top-level docs/, add profiling-framework.md and l2-swimlane-profiling.md, refresh tensor-dump.md, and update profiling-name-map.md / runtimes.md / testing.md to point at the new locations and the per-case output_prefix layout.

Testing

• Simulation tests pass
• Hardware tests pass

[gemini-code-assist]

Code Review

This pull request unifies the host-side infrastructure for PMU, L2 Swimlane, and Tensor Dump profiling into a shared framework, significantly reducing code duplication and improving maintainability across the a2a3 and a5 architectures. Key enhancements include a new three-bucket counter accounting model for better loss diagnostics, improved memory management via a centralized buffer pool, and the addition of completion barriers to ensure data consistency in tensor dumps. Feedback from the review suggests optimizing memory barriers in the SPSC queue logic to avoid redundancy and improve performance on weak-ordering architectures, as well as increasing the frequency of progress updates during the final data export phase.