Implement joint position limits

[diegoferigo]

This PR implements joint position limits by reading the following fields from a parsed SDF model:

• joint/axis/limit/stiffness
• joint/axis/limit/dissipation

Despite we read both parameters, the model used by the physics just uses the dissipation parameter as $k_{limit}$. This is because the stiffness/dissipation parameters could be mapped respectively to a spring/damper model, and the implemented model is equivalent to a pure spring. This means that the joint velocity is not used, and the torque is computed just as a term proportional to the joint position error wrt the limit.

I've tried also two different spring/damper models:

$$ \tau_{limit}^{(1)} = \begin{cases} k_{damper} (s_{min} - s) + k_{spring} \dot{s}, \quad \text{if } s < s_{min} \\ k_{damper} (s_{max} - s) - k_{spring} \dot{s}, \quad \text{if } s > s_{max} \\ \end{cases} $$

and

$$\tau_{limit}^{(2)} = \begin{cases} k_{damper} (s_{min} - s) + k_{spring} \text{abs}(s_{min} - s) \dot{s}, \quad \text{if } s < s_{min} \\\ k_{damper} (s_{max} - s) - k_{spring} \text{abs}(s_{max} - s) \dot{s}, \quad \text{if } s > s_{max} \end{cases}$$

Beyond requiring to tune an additional $k_{spring}$ parameter, both models do not give interesting additions. Particularly, $\tau_{limit}^{(1)}$ produces very often NaNs due to the discontinuous behavior (that would also affect differentiability #4). Instead, $\tau_{limit}^{(2)}$ enforces linearity by smoothing the spring gain with the position error, but beyond being having a less physical explanation, it works only with small $k_{spring}$ in order to prevent NaNs.

For the time being, I prefer keeping things simple. More advanced models can be introduced in the future. Note that a more advanced methodology exploits a constraint solver for joint limits, based on the Gauss principle of least constraint. Since it would make the forward step of the simulation much heavier, let's begin supporting the simple model of this PR.

Here below an example with a simple pendulum. GIFs are recorded with a real-time factor of 0.2.

w/o limits	w/ limits

Closes #20.