Reduce timer-enabled runtime overhead in direct .cod runs and cross benchmarks

[Copilot]

The issue was high execution overhead when running non-Invalid .cod tests directly (not parity) with timers enabled, and similarly inflated runtime in cross benchmarks under timer-enabled modes.
This change targets timer hot paths so profiling remains available while avoiding timer bookkeeping/log amplification in tight interpreter loops.

• Timer hot-path redesign (DebugSystem)

  • Replaced single-entry level timer maps with per-thread LIFO stacks keyed by operation name.
  • Supports nested same-name timers correctly without generating unique timer IDs per invocation.
  • Added explicit per-thread timer-state cleanup when debug level is turned off.

• Hot-path timer call simplification

  • Removed per-call timer-name construction (threadId + nanoTime) from perf helpers in:
    • ASTVisitor
    • ExecutionContext
    • ExpressionHandler
    • ImportResolver
    • NaturalArray
  • All now use stable operation names for start/stop in level-based timers.

• High-frequency timer level adjustment

  • Moved expression.handleBinaryOp perf timing from DEBUG to TRACE.
  • Keeps fine-grained visibility available when explicitly requested, while preventing debug-mode log/timer saturation in arithmetic-heavy code.

// Before
String timer = startPerfTimer(DebugSystem.Level.DEBUG, "expression.handleBinaryOp");

// After
String timer = startPerfTimer(DebugSystem.Level.TRACE, "expression.handleBinaryOp");