Because FactorVariable handles pointers (to its own member!) in its constructor, it needs a custom copy constructor and copy assignment operator.

To round out the rule of five (http://en.cppreference.com/w/cpp/language/rule_of_three) also define move constructor, move assignment, and destructor.

Add tests which would fail previously.

- Handle special case of size-1 variables

Variables of size 1 are a special case, allowing values to be given as doubles instead of Eigen::Matrix<double, 1, 1>.

- Add `inline` to FactorGraph method definitions

It is needed for ODR since the class is not a template

- Add tests

(Squashed commit)

- Add addPerfectPrior

- Simplify constructors

When we have to copy a value, accept a value.

The alternative is to have two overloads: (T&&) and (T const&) but this adds repeated code.

See https://stackoverflow.com/questions/7592630/is-pass-by-value-a-reasonable-default-in-c11

- Refactor perfect priors

- Setting variables constant is handled in graph_wrapper, not by factors
- Remove ZeroNoise type, instead, use "void" NoiseType
- evaluateRaw calls a different templated helper function if the Factor has zero noise

- Add PerfectPrior class

Remove the evaluateSpecialized helper, too complicated.

Have addPerfectPrior add a special PerfectPrior class. This class does nothing on evaluateRaw.

Set the variable value, as well setting it constant. This is the key!

Also:
- Fix immediate assignment of variable values

Also:
- Noise: pre-calculate inverse covariance, add docs

The solver relies on the reported jacobian being the actual Jacobian of what
it's given, not something scaled, so we have to multiply it as well.