I added some stopwatches to time main sections
of the plumed calculation part.
It might be interesting to include per-Action stopwatches
in the future.

I modified the command line tools stuff so that it is now possible
to access it using the Plumed.h interface (i.e. cmd(...)). This
allows to call the command line tools directly from other languages,
and is more consistent with the way plumed library is accessed from
MD codes. Notice that now also a plumed-runtime executable
is produced, which looks for PLUMED_KERNEL variable at runtime
to choose the proper library to load. Since all the command line
tools are built inside the library, this just means that a
different plumed version is used (similar to runtime mode
for patching MD codes).

Moreover, I hard-coded the PLUMED_ROOT variable in the
PlumedConfig.h header file. In this manner, it is sufficient to
have plumed in the path.

Basically I replaced all objects included in other objects with references, so as
to eliminate the need to include the corresponding header file. In this manner the
entanglement of compilation is greatly reduced.
Forward class declaration is now used everywhere excepted for:
* cases where it is impossible due to inline functions
  (for efficiency or because they are templates)
* cases where it is impossible due to inheritance
* small objects of fixed size such as Vector/Tensor/AtomNumber/Units/Pbc
* classes of std:: namespace (it is not possible to forward-declare them).

It takes care of dynamic loading, and unload libraries when plumedMain
is destroyed. This allows for a complete memory cleanup at the
end of execution.

I moved the tools to open a file with a specific prefix to PlumedMain.
This allows easily to use them also to open the input file, which in
turn means that we will be able to do bias-exchange and Hamiltonian
replica exchange

This is useful to compute acceptance.
I did it in the "dirty" way: if a value named "bias" exist inside any
action, it is accumulated into the total bias. I also changed the name
of values in Restraint and MovingRestraint so that the energy is now called
"bias" and non-capital letters are more consistent

It is now possible to automatically add a suffix to
all the files opened through the Action::fopen tool.
Opening is done as follows:
First the file+suffix is tried.
In case of error the file without suffix is tried.

In this manner, it should be possible to provide a single
input file also for replica exchange stuff. For output files,
it should make any difference.

I also added a stupid keyword to test this feature. Just use
_SET_SUFFIX pippo
in the plumed.dat file to add a suffix pippo to opened files

I am starting to merge my GREX branch, which should be
general enough to be used in any REX framework where
coordinates are swapped (a-la-gromacs). Since GREX
significantly modifies the MD-PLUMED interface through
the addition of several new cmd(), I tried to encapsulate
it as a separate objects with a cmd()-like interface.
Commands from MD starting with GREX are automatically
forwarded to that object. E.g.
plumed_cmd(plumed,"GREX init",NULL);
or so should initialize the GREX object, etc.

Up to this point I just upload the interface, later
on I will add the real GREX object, for which I still
have to do some cleaning.

In particular, I would like to be able to just do the
"forward" loop (calculate()) and not the "backward" loop
(apply()). In this manner, when computing foreign biases
for acceptance in replica exchange, I will only call the
forward loop, which computes colvars and biases. Clearly,
we should move printing/analysis/hills-adding stuff
outside the "calculate" method, so as to avoid doing these
things in the acceptance calculation. I wait for this further
cleaning since, until replica exchange is not actually there,
there is no practical difference. Also, we may decide to change
a bit the name of methods (calculate/apply are becoming a bit
misleading?)

I modified the "LOAD" command in such a way that if one
writes "LOAD pippo.cpp" in the plumed.dat file, pippo.cpp is automatically
compiled to a pippo.so which is then loaded.

Part of the interpretation of the cmd string can be moved here.
Also, it will be easy to define other classes with this interface
(I am planning to do that for gromacs-style replica exchange)

I removed several (hopefully) not needed include

I added some stricter check for deallocations of Atoms::Request
objects. I also removed the SIGMA keyword from Colvar (the right
place would be in Value). Finally, I fixed some lines which were
leading to apparent memory leakage with valgrind.

This is the draft version that I prepared and that we will use as
a starting point. Still it is a bit dirty, and I will try to improve
documentation and delete unused routines