Fix thread safety in fast_tanh toggle

[Ray0907]

---

Summary

• Fix thread safety issue in fast_tanh toggle by using std::atomic instead of modifying shared state
• Remove mutable global _activations map modification that could cause data races

Test plan

• Verify fast_tanh enable/disable works correctly in single-threaded context
• Test concurrent access from multiple threads toggling fast_tanh
• Confirm no performance regression with atomic load in hot path

---