This bug was introduced by me in 5e3aaf3a
The barrier on multi replicas was due to overcaution and
is only necessary when using WALKERS_MPI

It seems that sometime on travis a regtest with mw-mpi crashes.

I think this is because when using multiple processors per walker
there is a problem in the opening of the old file (for read) and
its backup (for write). Adding this barrier seems to make it more robust

Related to: e1b845b4

Now the type is defined with a typedef so that it is easier to change it.

I also set the type to size_t instead of unsigned long long.
This should be more portable.

Only add to the manual RESTART UPDATE_FROM and UPDATE_UNTIL
in the actions for which this makes sense

Allow individual action to set restart, using RESTART=YES or
RESTART=NO. By default, RESTART=AUTO, that means that global
restart settings are used.

Notice that this required to change the way OFile knows about
restart. Indeed, if the file is linked to an action,
it should use Action::getRestart instead of PlumedMain::getRestart.

Work is correctly computed for METAD and MOVINGRESTRAINT, and subtracted
to the effective energy drift

(cherry picked from commit b450f1525a218e8181f3933779bb17e550384f61)

Conflicts:
	src/bias/MetaD.cpp

I added the possibility to pass kbT from gromacs to plumed.
To allow compatibility between old and new versions, I used the API
number. Namely: gromacs checks that plumed API>1; if so, pass the
temperature.

KbT is only used in METAD so far (not in HISTOGRAM)

Addresses #35