Fix SLEEF preprocessor macro name to match ATen vec headers

The ATen NEON vectorized math headers (vec128_float_neon.h) check for
AT_BUILD_ARM_VEC256_WITH_SLEEF to enable SLEEF intrinsics for exp(),
log(), etc. ExecuTorch's get_vec_preprocessor_flags() was defining
ET_BUILD_ARM_VEC256_WITH_SLEEF (wrong prefix), so the USE_SLEEF macro
always took the fallback path: map(std::exp) — scalar exp called
per-element with full vector load/store overhead wrapping it.

With this fix, Vectorized<float>::exp() correctly dispatches to
Sleef_expf4_u10 on ARM, which is the intended behavior.

Differential Revision: [D96044314](https://our.internmc.facebook.com/intern/diff/D96044314/)

ghstack-source-id: 361224790
Pull Request resolved: #18645

Author: kimishpatel

kernels/optimized/lib_defs.bzl

@@ -25,16 +25,16 @@ def get_vec_preprocessor_flags():
         # various ovr_configs are not available in oss
         preprocessor_flags = select({
             "ovr_config//os:linux-x86_64": [
-                "-DET_BUILD_ARM_VEC256_WITH_SLEEF",
+                "-DAT_BUILD_ARM_VEC256_WITH_SLEEF",
             ] if not runtime.is_oss else [],
             "ovr_config//os:iphoneos-arm64": [
-                "-DET_BUILD_ARM_VEC256_WITH_SLEEF",
+                "-DAT_BUILD_ARM_VEC256_WITH_SLEEF",
             ] if not runtime.is_oss else [],
             "ovr_config//os:macos-arm64": [
-                "-DET_BUILD_ARM_VEC256_WITH_SLEEF",
+                "-DAT_BUILD_ARM_VEC256_WITH_SLEEF",
             ] if not runtime.is_oss else [],
             "ovr_config//os:android-arm64": [
-                "-DET_BUILD_ARM_VEC256_WITH_SLEEF",
+                "-DAT_BUILD_ARM_VEC256_WITH_SLEEF",
             ] if not runtime.is_oss else [],
             "DEFAULT": [],
         })