Use "-O3" for optimized TFLite build

This gives about 2x improvement with benchmark_model tool for Linux targets.
On Windows, the option is ignored since it isn't supported.

PiperOrigin-RevId: 278278192
Change-Id: I32433455db34b2dfe77266f9a9f705b54a66e96b

tensorflow/lite/build_def.bzl

@@ -13,12 +13,8 @@ def tflite_copts():
     copts = [
         "-DFARMHASH_NO_CXX_STRING",
     ] + select({
-        clean_dep("//tensorflow:android_arm64"): [
-            "-O3",
-        ],
         clean_dep("//tensorflow:android_arm"): [
             "-mfpu=neon",
-            "-O3",
         ],
         clean_dep("//tensorflow:ios_x86_64"): [
             "-msse4.1",
@@ -30,6 +26,9 @@ def tflite_copts():
         "//conditions:default": [
             "-Wno-sign-compare",
         ],
+    }) + select({
+        clean_dep("//tensorflow:optimized"): ["-O3"],
+        "//conditions:default": [],
     })
 
     return copts