Make TokenInfo use a single 32-bit store for initialization. (#8054)

We need to zero-initialize the `token_payload_` bits in order for the
normal creation to just be zeroing the whole structure. We do this only
when MSan isn't enabled so we don't lose bit-accurate bug detection.

Assisted-by: Antigravity with Gemini

toolchain/lex/token_info.h

@@ -162,14 +162,19 @@ class TokenInfo {
   //
   // Only used by the lexer which enforces only the correct kinds are used.
   //
-  // When the payload is not being set, we leave it uninitialized. At least in
-  // some cases, this will allow MSan to correctly detect erroneous attempts
-  // to access the payload, as it works to track uninitialized memory
-  // bit-for-bit specifically to handle complex cases like bitfields.
+  // Under MSan, leave `token_payload_` uninitialized so bit-level tracking can
+  // detect erroneous attempts to access a missing payload. In non-MSan builds,
+  // zero-initialize it so constructing a payloadless `TokenInfo` compiles to a
+  // plain 32-bit store rather than a read-modify-write preserving uninitialized
+  // upper bits (which SROA otherwise promotes to loop-carried registers).
   TokenInfo(TokenKind kind, bool has_leading_space, int32_t byte_offset)
       : kind_(kind),
         has_leading_space_(has_leading_space),
-        byte_offset_(byte_offset) {}
+#if !__has_feature(memory_sanitizer)
+        token_payload_(0),
+#endif
+        byte_offset_(byte_offset) {
+  }
 
   // Constructor for a TokenKind that carries a payload.
   //