Refactor: separate orch phase-stats and swim-lane gating

[ChaoWao]

Summary

Split the two profiling concerns in the PTO2_ORCH_PROFILING branch of
CYCLE_COUNT_LAP_RECORD so each is gated independently:

• Phase-statistics accumulation (acc += _t1 - _t0) — gated only by
  PTO2_ORCH_PROFILING at compile time, unconditional at runtime.
• Swim-lane GM writes — gated only by orch->l2_perf_level >= L2PerfLevel::ORCH_PHASES at runtime, mirroring the
  PTO2_PROFILING branch and the matching reads in
  l2_perf_collector.cpp / aicpu_executor.cpp.

Why

Previously the PTO2_ORCH_PROFILING branch bundled cycle accumulation
and swim-lane writes behind a single compile-time gate, with two
consequences:

1. Callers could not get phase totals without paying GM-store cost on
   every phase boundary.
2. The swim-lane write happened between the _t1 capture and the
   _t0 reassignment, so its cost leaked into the next phase's
   accumulator and distorted the totals the flag exists to measure.

After this change the swim-lane write is followed by a fresh
get_sys_cnt_aicpu() for _t0, so its GM-store cost is excluded from
the next phase's accumulator.

Mirrored across a2a3 and a5 orchestrator implementations.

Testing

• Default sim build (compile-time gate stays off — no behavior change)
• PTO2_ORCH_PROFILING=1 build on hardware: phase totals stable
  across runs regardless of --enable-l2-swimlane level
• --enable-l2-swimlane 4 with PTO2_ORCH_PROFILING=1: swim-lane
  records present, totals not inflated by GM-store cost

[gemini-code-assist]

Code Review

This pull request updates the profiling macros in pto_orchestrator.cpp across both the a2a3 and a5 runtimes to conditionally gate swim-lane recording at runtime based on l2_perf_level and re-sample the cycle counter after recording to exclude write overhead. The review feedback points out a fragile implicit dependency in the CYCLE_COUNT_START macro, which relies on a local variable named orch being present in the scope. It is recommended to use this->l2_perf_level instead to make the macro self-contained and prevent potential compilation errors.