Improve the `build` docstrings about lampe

[francois-rozet]

Would the following be enough?

build: The network constructor (e.g. :class:`lampe.nn.ResMLP`). It takes the
    number of input and output features as positional arguments.

For the embedding tutorial, the following paragraph explains why we need an embedding. Should we add something?

Because the observations are images, they cannot be fed directly to a NPE module, which only accepts vector-shaped observations. Instead, we rely on a convolutional neural network to extract a vector of informative features from the observation and feed this vector to the density estimator. In our case, the latter is a neural spline flow (NSF) with rational-quadratic splines borrowed from the zuko package.

[francois-rozet]

The constructor descriptions have been updated in 83fd155.