Gaussian kernel functions were normalised incorrectly.  This should
only cause a multiplicative scaling of the histograms calculated
using the HISTOGRAM action

If you had an action that inherited from both ReferenceArguments and
ReferenceAtoms then there was a variable that appeared with the same name
in both so you can get conflicts.  In addition, RMSD object in optimal RMSD
needs clearing prior to setting.  Lastly made it so that the metric is printed
out in emdbedding files.

changelog, regtest and bug fixes

close #172

related to #137

The kernels were normalised incorrectly and their derivative when the
covariance matrix was non-diagonal was computed wrongly.  This check
affects a number of regtests.  analysis/rt0 changes beecause the kernels
were normalised incorrectly.  basic/rt44, basic/rt48, basic/rt49, basic/rt52
change because the derivatives output in the fes file have changed sign

Before this fix the phase was chosen enforcing the sum of the elements
of each eigenvector to be positive (see dd4bd09b).
This does not work well when the
sum is numerically close to zero for some symmetry. With this fix I
set the first non-zero element of each eigenvector to be positive.
Hopefully it is more robust and make results better reproducible
for regtesting

This test is very sensitive to numerical errors and is thus
not a good test of the code.  The functionality that is
being tested here is tested elsewhere.

This regtest always fails because it was not very well designed.
Furthermore, the functionality being tested by this test is checked in analysis/rt0
so keeping this bad test seems not worthwhile.

I am not 100% sure this will make it reproducible on different machines
(maybe we are trying to fight Lyapunov). Anyway, for sure this
test was not reset after commit 77aa3cd9
and this was a mistake

I added a few lines to set to a predictably value the
arbitrary phase in the eigenvectors used for MDS.
This should make the runs more reproducible.

Problem due to very large values for cell vectors with nopbc

Also fixed how weights are dealt with in analysis and mapping
weights were stored in different places in analysis and mapping
this was confusing for developers.  The weights are now always
inside the individual ReferenceConfiguration objects

Regtest crashes

(should be checked by GT)

This test shows a problem in reweighting (discussed in #48).
In particular, histoA and histoB in this regtest should not
come out identical as they do.

This allows to run regtests inside e.g. regtest/basic directory

It was a longstanding (unknown) bug!
I added reference output files for all the simpledmd
regtest. This will allow to distinguish bugs in plumed
from Lyapunov effects in trajectories.

Made it so that virtual routine is not called from constructor.  If the calculation
is terminated during the analysis phase then the analysis is now restarted when
the calculation reaches update on the first loop.  I also removed the duplicated
backup code that was in analysis.  Backup is now done by the routines in OFile.
This routine is now a little more flexible in that you can specify the string
in the backup name.