Call tzset just once, at the start, rather than for every test
case. This lets us measure the CPU cost of mktime as opposed to
that of tzset. This is relevant when TZ is not set and glibc is
being used. This speeds up tests by a factor of 40 on my Fedora
23 x86-64 platform.
* lib/mktime.c (main) [DEBUG_MKTIME]: Call localtime at the start,
to call tzset and as a sanity check. Later on, use localtime_r
instead of localtime.