Avoid zero-initializing our most-used buffers.

The StringPrintf one is heavily used and brings the overhead versus
fmtlib down to 1.5x rather than 2x. I don't have a convenient benchmark
for the other two.

Test: libbase tests & benchmarks
Bug: http://b/155324241
Change-Id: I9e704a360846d5520c53f668e7c315b0c0ea55f8

base/file.cpp

@@ -225,7 +225,7 @@ bool ReadFdToString(borrowed_fd fd, std::string* content) {
     content->reserve(sb.st_size);
   }
 
-  char buf[BUFSIZ];
+  char buf[BUFSIZ] __attribute__((__uninitialized__));
   ssize_t n;
   while ((n = TEMP_FAILURE_RETRY(read(fd.get(), &buf[0], sizeof(buf)))) > 0) {
     content->append(buf, n);

base/logging.cpp

@@ -260,7 +260,7 @@ static void KernelLogLine(const char* msg, int length, android::base::LogSeverit
   // The kernel's printk buffer is only 1024 bytes.
   // TODO: should we automatically break up long lines into multiple lines?
   // Or we could log but with something like "..." at the end?
-  char buf[1024];
+  char buf[1024] __attribute__((__uninitialized__));
   size_t size = snprintf(buf, sizeof(buf), "<%d>%s: %.*s\n", level, tag, length, msg);
   if (size > sizeof(buf)) {
     size = snprintf(buf, sizeof(buf), "<%d>%s: %zu-byte message too long for printk\n",

base/stringprintf.cpp

@@ -25,7 +25,7 @@ namespace base {
 
 void StringAppendV(std::string* dst, const char* format, va_list ap) {
   // First try with a small fixed size buffer
-  char space[1024];
+  char space[1024] __attribute__((__uninitialized__));
 
   // It's possible for methods that use a va_list to invalidate
   // the data in it upon use.  The fix is to make a copy