aes_gcm: Prefer integrated implementations, avoid inlining fallbacks.

The `#[inline(always)]` on `open_strided` was just a mistake.

But, even correcting that, the compiler won't structure `open()` and
`seal()` in the expected way. With these changes, the integrated
implementations will get inlined (as far as possible) and the
compiler will be biased towards assuming they will be used.

src/aead/aes_gcm.rs

@@ -218,6 +218,26 @@ pub(super) fn seal(
     }
 }
 
+#[cfg_attr(
+    any(
+        all(
+            any(target_arch = "aarch64", target_arch = "arm"),
+            target_feature = "neon"
+        ),
+        all(
+            any(target_arch = "x86", target_arch = "x86_64"),
+            target_feature = "sse"
+        )
+    ),
+    inline(never)
+)]
+#[cfg_attr(
+    any(
+        all(target_arch = "aarch64", target_feature = "neon"),
+        all(target_arch = "x86_64", target_feature = "sse")
+    ),
+    cold
+)]
 fn seal_strided<A: aes::EncryptBlock + aes::EncryptCtr32, G: gcm::UpdateBlocks + gcm::Gmult>(
     Combo { aes_key, gcm_key }: &Combo<A, G>,
     aad: Aad<&[u8]>,
@@ -392,7 +412,26 @@ pub(super) fn open(
     }
 }
 
-#[inline(always)]
+#[cfg_attr(
+    any(
+        all(
+            any(target_arch = "aarch64", target_arch = "arm"),
+            target_feature = "neon"
+        ),
+        all(
+            any(target_arch = "x86", target_arch = "x86_64"),
+            target_feature = "sse"
+        )
+    ),
+    inline(never)
+)]
+#[cfg_attr(
+    any(
+        all(target_arch = "aarch64", target_feature = "neon"),
+        all(target_arch = "x86_64", target_feature = "sse")
+    ),
+    cold
+)]
 fn open_strided<A: aes::EncryptBlock + aes::EncryptCtr32, G: gcm::UpdateBlocks + gcm::Gmult>(
     Combo { aes_key, gcm_key }: &Combo<A, G>,
     aad: Aad<&[u8]>,