More robust fitting for LC reduction

Add some more levels of reduction. This would generate a "smoothed flux" and "detrended flux" to be used downstream in the analysis.

This should be observatory-specific, hence in the reducer.

For ZTF this could include just removing a 3rd order polynomial from each filter/season.

For TESS this could include:
- getting rid of NaNs in a smart way (think of edge effects for filters)
- smoothing using a sine/cosine fit and then a Savitzky-Golay filter