Added evaluation of expressions when inserting in Values

[arutkowski]

This PR removes the inconvenient need to evaluate expressions involving values of type Point2 or Point3 when passing them to Values::insert, as mentioned in #1489.

In other words, using Class Template Argument Deduction (CTAD) a user can now simply call insert(key, a + b) instead of insert(key, (a + b).eval()) or insert<Point2>(key, a + b).

The related functions Values::insert_or_assign and Values::update have been modified in a similar manner.

Unit tests have also been added to verify the functionality.

---

Details

The inconvenience was originally caused by the fact that Point2 and Point3 are really just Eigen::Vector2 and Eigen::Vector3, and since Eigen uses lazy evaluation, expressions involving vectors are not automatically evaluated to vectors until they are needed (which is why the use of the auto keyword is discouraged in many cases).

The solution that was implemented here was to add template specializations for values of type Eigen::CwiseUnaryOp and Eigen::CwiseBinaryOp. Both are needed so expressions can be chained together, as in 2*a + b - c.

Following the guidance for designing functions that take Eigen types as parameters, I tried to implement the following function prototype:

    template <typename Derived>
    void insert(Key j, const Eigen::MatrixBase<Derived>& val);

However, this produces a lot of errors related to invalid use of incomplete type struct Eigen::internal::traits<gtsam::Rot3>, which I don't currently understand.

Related Future Work

There is a related inconvenience involving the creation of factors, as in BetweenFactor(key1, key2, b - a, noise). It would be a bit more difficult to implement implicit evaluation for Eigen types of a and b since the base class of NoiseModelFactorN used for most factors only includes the noise model and not the actual measurement (b - a in this case). Thus, every factor would need to implement template specializations for Eigen::CwiseUnaryOp and Eigen::CwiseBinaryOp, or perhaps they could all inherit from something like a GenericFactor that would include both the noise model and the measurement.

[varunagrawal]

LGTM