mktime: fix some integer overflow issues and sidestep the rest

mktime: fix some integer overflow issues and sidestep the rest

This was prompted by a bug report by Benjamin Lindner for MinGW
<http://lists.gnu.org/archive/html/bug-gnulib/2011-01/msg00472.html>.
His bug is due to signed integer overflow (0 - INT_MIN), and I
I scanned through mktime.c looking for other integer overflow
problems, fixing all the bugs I found.

Although the C Standard says the resulting code is still not safe
in the presence of integer overflow, in practice it should be good
enough for all real-world two's-complement implementations, except
for debugging environments that deliberately trap on integer
overflow (e.g., gcc -ftrapv).

* lib/mktime.c (WRAPV): New macro.
(SHR): Also check that long_int and time_t shift right in the
usual way, before using the fast-but-unportable method.
(TYPE_ONES_COMPLEMENT, TYPE_SIGNED_MAGNITUDE): Remove, no longer
used.  The code already assumed two's complement, so there's
no need to test for alternatives.  All uses removed.
(TYPE_MAXIMUM): Don't rely here on overflow behavior not defined by
the C standard.  Problem reported by Rich Felker in
<http://lists.gnu.org/archive/html/bug-gnulib/2011-01/msg00488.html>.
(twos_complement_arithmetic): Also check long_int and time_t.
(time_t_avg, time_t_add_ok, time_t_int_add_ok): New functions.
(guess_time_tm, ranged_convert, __mktime_internal): Use them.
(__mktime_internal): Avoid integer overflow with unary subtraction
in two instances where -1 - X is an adequate replacement for -X,
since the calculations are approximate.