Cleaned up run time measurement

Cleaned up run time measurement

Removed all use of clock(), in favour of using clock_gettime for
per-node timer and (where available) per-thread timer. This is better
than using gettimeofday(), which is deprecated by POSIX. clock() has
no real advantages, and requires us to keep bumping our own counter if
we use it (in particular, on any remaining 32-bit
machines). Eliminated related dead code.

Cleaned up struct gmx_runtime_t and related code by
* renaming variables more descriptively
* removing variables that should be local variables in a function
* eliminating variables associated with using clock()

The behaviour of print_perf is preserved (tested real MPI, thread MPI;
serial looks like it should work also); namely that the "(%)" should
be a crude reflection of machine load.

Change-Id: Icdf506150045bee88cf04200b25cd010a9bb5419