`Constraint` differences from the base class

[willGraham01]

Constraints differ from CausalEstimands because they will also come with an associated tolerance, as well as "data value".

Their general structure can be thought of as:

$$ \mathrm{norm}\left( g(\theta) - g_{data} \right) - \epsilon \leq 0. $$

• $g(\theta)$ is the quantity as estimated from the model, which is a function provided by the user.
• $g_{data}$ will be the data provided by the user. This may change between runs using the same model structure.
• Constraints are all scalar valued, so even though $g$ and $g_{data}$ are vectors, $\mathrm{norm}(\cdot)$ is needed to take a suitable norm.
• $\epsilon$ is a tolerance provided by the user. Again, this may change between runs.

As such, Constraints are better defined in stages:

• The user provides $g(\theta)$ at initialization. This is a hidden attribute and isn't meant to be changed.
• The user also provides the norm function at initialization. Again, the intention is that this does not change. It should default to the L^2 norm (which is the absolute value in 1D).
• $g_{data}$ and $\epsilon$ are public attributes that can be changed, but also set at initialization.
• Constraint.__call__ needs to be overwritten to return the appropriate composition of these attributes.