Option to use CBVs and better priors

[jorgemarpa]

Two small changes/new features:

• machine.build_time_model() accepts other_vectors as argument to add other type of model components from user-level, e.g. using CBVs as regressors.
• Improved priors for in PerturbationMatrix. The prior follows ~ N(1, 0,5) which enforces "small" deviations from the mean model. This accounts for the mismatch in mean flux we saw before between PSF and PSF-NOVA and reduces the chances to get outlier data points.

TODO:

• Fix test_perturbation_matrix3d()

Requires #60 to be merged first.

[jorgemarpa]

Using smooth CBVs instead of PCA components and adjusting the priors for the perturbation model improved the deviation from the mean_model as seen in the following figures:

x-axis: sum of the mean_model in the pixel dimension, i.e. the total PSF per source
y-axis: time-wise mean and standard deviation of the summed perturbed_model per source
histogram: ratio of above values

Data points of the 1-1 relation are sources with mismatching mean light curves before and after the time model and sources with spiky features (or added variability) after applying the time model.

After narrowing the priors the sources with bad time models are fixed.

[christinahedges]

This is 99.999%, I will merge when @jorgemarpa checks that the segments keyword is the desired behavior.