Using c++11 to have constructors of generic size.

Two types of optimization implemented:

1. Tensor is used instead of Matrix. This allows avoiding the extra
allocations required in Matrix to setup the workspace, since the
needed space is now on the stack rather than on the heap

2. When align==displace, only lower eigenvalue is computed.

Overall, these changes result in a 5-10% speedup of rt46 on my mac

I used Tensor/Vector operations now

Notice that one can provide an evec matrix with less rows.
In this case, the lower eigenvalues will be found. In addition,
temporary data is on the stack (array) rather than on the heap (vector),
making it faster for small matrices.

see 7a72bc37

Using completion the config.sh script is called during compilation.
This fix avoids problems when cltools is compiled before config

One could just use a plain object in these cases (no pointer)
but I did the minimal change to remove the delete statement.

This reverts commit d190c105.

This reverts commit 40099475.

This reverts commit 14eb6054.

This reverts commit 8e6a72aa.

Previous import was dating back to gromacs 5.1.5
This brings in a few changes, notably:

- source has been updated to c++
(see https://github.com/gromacs/gromacs/commit/eb81388df76ddb91e370a4524d4226697ea3a3ef)

Notice that some of the changes require c++11 and thus should not be backported to
PLUMED < 2.4

Previous import was dating back to gromacs 4.6.7.
This brings in a few changes, notably:

- source tree was modified

- some thresholds were modified
(see https://github.com/gromacs/gromacs/commit/97161d2601f4f98a45d49de2ef8f630f18c94455)