relax SVN archive

Added the 'NS 2-site star red' model to the specific_analyses.relax_disp.variables module.

This is the model of the numerical solution for the 2-site Bloch-McConnell equations using complex
conjugate matrices, whereby the simplification R20A = R20B is assumed.

This commit follows step 1 of the relaxation dispersion model addition tutorial
(http://thread.gmane.org/gmane.science.nmr.relax.devel/3907).

Reworked the dispersion chapter of the manual for the recent support of numerical models.

This includes better sectioning and section labelling and referencing, and the addition of the
'NS 2-site star' numerical model. The model and parameter tables have been updated as well.

Changed 'numerical integration' to 'numerical solutions' in the dispersion chapter of the manual.

Docstring completion for lib.dispersion.ns_2site_star.r2eff_ns_2site_star().

Epydoc text was missing for some of the function arguments.

Improvement of the error handling in the 'NS 2-site star' model.

The fA and pB parameters are no longer being checked. Instead a Mgx value of 0.0 is being checked
for. This catches additional problems. And now instead of the R2eff value being set to zero, it is
set to 1e99. This is because log of zero is -inf, and then multiplied by a negative constant gives
positive inf.

More speed ups of the 'NS 2-site star' dispersion model.

A number of calculations have been shifted to the target function initialisation code, avoiding
unnecessary repetitive mathematical operations.

Shifted to using the faster numpy.linalg.matrix_power() function in lib.dispersion.ns_2site_star.

This was originally using the lib.linear_algebra.matrix_power.square_matrix_power() function,
however the numpy equivalent is faster.

Speed ups of the 'NS 2-site star' dispersion model optimisation.

The relaxation and magnitisation data structures are now initialised with the target function
initialisation, rather than being created at each target function call. The Rex and M0 matrices are
now updated at the base of the target function rather than in the lib.dispersion.ns_2site_star
module to minimise the number of mathematical operations per target function call. And the M0
matrix has changed shape and a dot product is used in lib.dispersion.ns_2site_star to create Moft
instead.

The 'NS 2-site star' model is now more robust against math domain failures.

This includes the failure of the logarithmic of zero matrices.

Fix for the state G+E to A+B conversion in lib.dispersion.ns_2site_star.