[[Imported from SVN: r9083]]

Not tested yet.

[[Imported from SVN: r9082]]

[[Imported from SVN: r9081]]

[[Imported from SVN: r9080]]

[[Imported from SVN: r9079]]

[[Imported from SVN: r9078]]

This is just an initial commit.  It is a copy of the corresponding code for
unit vectors, with just a few names exchanged.  The actual functionality
will come in subsequent commits.

[[Imported from SVN: r9076]]

[[Imported from SVN: r9023]]

[[Imported from SVN: r8983]]

[[Imported from SVN: r8982]]

[[Imported from SVN: r8981]]

[[Imported from SVN: r8974]]

[[Imported from SVN: r8972]]

[[Imported from SVN: r8862]]

[[Imported from SVN: r8861]]

[[Imported from SVN: r8687]]

[[Imported from SVN: r8630]]

[[Imported from SVN: r8629]]

[[Imported from SVN: r8627]]

There were two issues fixed by this patch:

a) The bending energy only involves the third director.
   One must look at it as a function in S^2, not as
   a subset of coefficients of a function to SO(3).
   This means that we have to use a different projection
   in one place.
b) A certain subset of the coefficients (see the code)
   must always be zero, because they are parts of
   3-vectors which are known to be perpendicular to (0,0,1).
   For some strange reason they aren't.  I now hardwire
   them to be 0 anyway.  This makes my test pass.
   Maybe one day I'll find the true reason :-)

[[Imported from SVN: r8626]]

of Neumann boundary forces.  But I am not sure and I want to
postpone the issue.  Therefore this patch adds a cpp warning.

[[Imported from SVN: r8625]]

Don't use a loop with 4 iteration to clear an array,
if the array has only length 3.

Boy, this did take a while to find...

[[Imported from SVN: r8593]]

Another easy patch which makes the code a lot faster.

Also, this patch removes the quasi-Newton code.  So far that code
hasn't helped at all.  Quasi-Newton methods will need more work
in the future.

[[Imported from SVN: r8588]]

the ones with i<=j.  The rest follows from symmetry of
the Hessian.

Needless to say, this makes the Hessian assembly quite
a bit faster.

[[Imported from SVN: r8587]]

[[Imported from SVN: r8572]]

[[Imported from SVN: r8571]]

[[Imported from SVN: r8570]]

[[Imported from SVN: r8569]]

Because trace(sym(A)) == trace(A)

[[Imported from SVN: r8568]]

[[Imported from SVN: r8567]]

[[Imported from SVN: r8566]]

[[Imported from SVN: r8565]]

[[Imported from SVN: r8562]]

[[Imported from SVN: r8559]]

[[Imported from SVN: r8554]]

[[Imported from SVN: r8553]]

[[Imported from SVN: r8543]]

[[Imported from SVN: r8508]]

[[Imported from SVN: r8507]]

This change makes the localgeodesicfefunction test fail again.
So there must be a bug in the corresponding non-finite-difference
code.  Looking at simplex cases confirms me in this opinion.
But I don't find the bug...

[[Imported from SVN: r8463]]