Discrepencies in conventions between planar and free joints

[aescande]

In Featherstone's book, there is two competing conventions for representing the translation part in planar joints and free joints.
Fig. 4.9 and table 4.1 (pp. 78-79) consider that the translations are expressed in successor frame (FS), but example 4.5 (p. 81) makes the case for having (translation) velocity variables expressed in FS but position variables expressed in parent frame (FP). This is to get better numerical stability while keeping simple motion subspace matrices S.

From the code of Joint::pose, planar joints are using the convention of Fig. 4.9, while free joints use that of example 4.5. I would push for having better coherency and change the convention of planar joint to the latter option: we would gain not only on the numerical behavior front, but also it is also, imho, more intuitive for the user to express the translation independently of the rotation.

In itself the change is drastic as it can break client code. In practice, it seems that the planar joint is not used much.
I changed locally Joint::pose for planar joints, and it didn't break any test. It seems to me that this is the only place the code need to change, but I don't know the whole library well enough, so there might be other places to amend. Maybe @jorisv has some advice on that.

It would be better to take a decision before addressing #35 to avoid additional work.