SDPA: use exp_u20 for the softmax exponential (pytorch#22082)

Summary:

Switch the softmax exponential in the flash-attention custom SDPA
kernel (`op_sdpa_impl.h`) from `Vectorized::exp()` to
`Vectorized::exp_u20()`. The file already carried this as a TODO.

This is not bit-exact. `exp_u20` is a ULP-20 approximation, so it differs
from `exp` outright, and fourteen autoregressive layers on that is enough
to flip a near-tie argmax.

Reviewed By: JakeStevens

Differential Revision: D117198987

Pull Request resolved: pytorch#22082

Author: pssrawat

extension/llm/custom_ops/op_sdpa_impl.h

@@ -593,9 +593,7 @@ _exp_reduce_sum_fusion_kernel(T1* a, const int& size, T2* out, T1& val) {
   for (int i = 0; i < vec_size * (size / vec_size); i += vec_size) {
     auto tmp0 = vec::VectorizedN<T1, 2>::loadu(a + i);
     auto tmp1 = tmp0 - vec_max;
-    // Replace with exp_u20 later
-    // auto tmp2 = tmp1.exp_u20();
-    auto tmp2 = tmp1.exp();
+    auto tmp2 = tmp1.exp_u20();
     vec_tmp_sum = vec_tmp_sum + tmp2;
     tmp2.store(out + i);
   }