reverting attn exp calculations to before 3n

[ncylich]

You can see there's no more diffs between this and the prior to 3n commit's kernel_attention.cpp:
d5cd5c0...d0ad24a

[Copilot]

Pull request overview

This PR reverts the vectorized exp/exp2 approximation used inside the FP16 attention softmax path back to the pre-“3n” polynomial form, aligning kernel_attention.cpp with the earlier implementation referenced in the PR description.

Changes:

• Replaced the higher-order fused-multiply-add polynomial for the fractional exponent with a 2nd-order Taylor approximation using ln(2) and ln(2)^2/2.
• Applied the same change in both the Apple Accelerate-based attention path and the general NEON vectorized attention path.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.