Fix signal race in iread. (#280)

The _setjmp call may be half-through with its operation when it is
interrupted by _longjmp. This can happen if iread() reaches line
172 which calls getchr, which in itself can call iread() again.

Effectively, it means that "reading" is already 1 when SET_JUMP is
called. A badly timed SIGINT could therefore mix up register states
and stack pointers of two different situations.

It can be reliably reproduced by using tmux and entering
":send-keys F C-x" to quickly send F and CTRL+x to less.

My tests on Linux/amd64 did not reveal a way to crash less or do
worse stuff, but let's be safe than sorry.

The fix is quite simple: Do not call SET_JUMP from within nested
iread calls. If reading is already 1, keep the read_label as it is.

os.c

@@ -119,7 +119,7 @@ iread(fd, buf, len)
 	}
 #endif
 #endif
-	if (SET_JUMP(read_label))
+	if (!reading && SET_JUMP(read_label))
 	{
 		/*
 		 * We jumped here from intread.