When the training data contains NaNs, Prophet's preprocess method panics:

panicked at /augurs/crates/augurs-prophet/src/prophet/prep.rs:318:40:
index out of bounds: the len is 297 but the index is 297

This is because we filter ds and y to remove NaNs, but don't adjust n accordingly (or any of the other seasonality/regressor fields, while we're at it).

We should move this filtering into a TrainingData::filter_nans() method. Ideally it would even consume self and return a new TrainingData, which would mean we could avoid having a mutable binding to data in preprocess and be sure we're not inadvertently mutating it anywhere else.