Use buffered stdio

Our generation of HTML triggers many small write(2) syscalls
which is inefficient.

Time output on a horrible query against my git.git mirror
shows significant performance improvement:

QUERY_STRING='id=2b93bfac0f5bcabbf60f174f4e7bfa9e318e64d5&id2=d6da71a9d16b8cf27f9d8f90692d3625c849cbc8'
PATH_INFO=/mirrors/git.git/diff
export QUERY_STRING PATH_INFO
time ./cgit >/dev/null

Before:
real    0m1.585s
user    0m0.904s
sys     0m0.658s

After:
real    0m0.750s
user    0m0.666s
sys     0m0.076s

1 files changed, 7 insertions, 0 deletions

diff --git a/cache.c b/cache.c

index 1c843ba..5937254 100644
--- a/cache.c
+++ b/cache.c
@@ -270,6 +270,13 @@ static int process_slot(struct cache_slot *slot)
 {
 	int err;
 
+	/*
+	 * Make sure any buffered data is flushed before we redirect,
+	 * do sendfile(2) or write(2)
+	 */
+	if (fflush(stdout))
+		return errno;
+
 	err = open_slot(slot);
 	if (!err && slot->match) {
 		if (is_expired(slot)) {