I implemented a framework to provide the virial corrections for
virtual atoms that are non-linear functions

I added the possibility to dump the 9 components of the virial
from the driver.

This is not strictly necessary since virial is
set anyway (and not accumulated), but I added
it for consistency.

If you had your atoms distributed among the domains such that one
node had on all the Atoms there could be a problem in running plumed.
This change fixes this problem.  Also deleted rogue check in
AdjacencyMatrixAction

[makedoc]

I changed a few things:
1. instead of checking with "--enable-debug" and "--enable-debug --enable-debug-glibcxx"
  I always use both.
2. I check with debug flags and serial compiler (before it was only debug+MPI)
3. I changed the order so that runs with manual is the first and runs
  with debug options are also made earlier. This is convenient since they
  take more time

This allows reading crd files without linking external plugins

(goes with 32c10c65)

ReferenceConfiguration::getReferencePositions() was hidden by
ReferenceAtoms::getReferencePositions()
since the former was const. I changed the former to non const.

@gtribello might want to check this

Multiple processors per replica. Forces are also checked

A collective variable value was passed instead of its index.
Found thanks to bound checking!

@carlocamilloni you might want to double check

If they are not too slow, they could be useful for finding bugs

[makedoc]

[makedoc]

[makedoc]

Notice that these tests will not run on travis-ci
since VMD plugins are not available there

This allow amber formats to be properly read

[makedoc]

This bug was not detected since it only appears if the reference pdb
has center not in the origin. I also added an extra regtest to check for
this

I changed modifyForce to modifyGlobalForce. Notice that other modifyXX
methods in ActionAtomistic are returning the local copy. "Global"
makes more clear that one is modifying the global copy of forces.

I also added modifyGlobalVirial with an analogous functionality

Fixes a potential problem in the following commits:
423d49bf
bb6178a2
(reported by C. Perego)

[makedoc]

[makedoc]

It does not shift anymore the atoms to their original positions.

Notice that this makes its behavior consistent with WHOLEMOLECULES
where atoms are not restored in the backward loop.

This implies that (consistently) one cannot rely that shared atoms
(those accessed with modifyAtoms()) are in their original position
in the ::apply() methods.

Same as 423d49bf applied to gmx 5.0

Needs gromacs to be repatched

Workaround for #132.

Notice that gromacs wants something like -0.5*f*x,
whereas plumed provides something like -f*x.

Additionally, for not so clear reasons, when using plumed without computing
ENERGY the virial contribution from plumed was completely ignored.

There are two workarounds in this patch:
1. In case plumed does not ask gromacs to compute ENERGY, the plumed contribution
   to virial is stored on a separate tensor and added to the gromacs one at the end
2. In case plumed asks gromacs to compute ENERGY, the a temporary tensor
   is created containing twice the gromacs virial and passed to plumed. The virial
   is then set to half of it.

This should solve both the issues discussed in #132

Needs gromacs to be repatched.