As of this commit, to patch with Hamiltonian replica exchange
just use the gromacs-4.6.3 patch

This will simplify keeping up to date with upstream changes in the patch.

I added a new mode to store patches. If one saves them with
--save-originals (instead of -s) what happens is that
/patches/codename.diff
will be a directory. Instead of storing diff files there,
we will store original/modified pairs.

When patching, if a directory is found in
/patches/codename.diff
the diff are computed and then reapplied on the actual files.

The only change is that the diff is not stored but re-computed on-the-fly.
This has a completely equivalent effect of the original way.
E.g., one can patch two slightly different versions
of an MD code with the same patch.

However, it has the advantage that stored files are more friendly.
E.g., one could directly edit them. Also the git-merge of two
different patches should work better now.

To be consistent with the paper, plumed is always capital letters.
I also changed 2.0 -> 2 and 1.0 -> 1 for sake of generality

Without STORE_GRIDS it is necessary to flush the grid file,
otherwise it might stay empty for the entire simulation

When not RESTARTing:
file is backed up if existing.
  with STORE_GRIDS, file grows during simulation
  without STORE_GRIDS, file is replaced every time grid is rewritten

When RESTARTing:
  with STORE_GRIDS, file is appended and grows during simulation
  without STORE_GRIDS, file is replaced every time grid is rewritten

This hopefully fixes #34
  

It is possible to print html links to failed tests.
I am setting a nightly build and want to publish online
regtests results

(pointed out by Xiaodong Pang)

One can know use e.g.
./configure mpi.linux.icc
to directly select the configuration and avoid interactive
choice. Better for building from script

Relevant when using a huge number of restraints

Running a LJ in LAMMPS (no charges) was crashing

cmd("useNaturalUnits") was erroneously crashing

Removed many scans of the output buffer by keeping trace of buffer length.
Time needed to write a long line with a sequence of file.printf() is
now linear with line length. Before it was quatradic, causing problems
especially when logging huge atom lists at startup.

I added a check that disables this combination which can
be problematic. I had to also update a regtest

You can now read in groups with names other than GROUPA, GROUPB and
GROUPC.  Also output of the atoms in groups is neater.  Lastly I realized that I
had accidentally added some crap to regtest/multicovar/rt45 - I have
now deleted this.

* Better algorithm for int powers
* Simplification when 2*N=M (which is the usual case)

(kindly provided by Xiaodong Pang)

There was an error in dihedral scaling (they were scaled with
sqrt(lambda) also if both terminal atoms were in the hot region)