Spike - Strike balance between speed and accuracy for ARIMA

I'd like ARIMA to use covariates as often as it can, as this leads to an increase in predictive accuracy. However with larger datasets, using covariates leads to exponentially larger fit times.

This issue tracks:

- setting covariate to False for ARIMA as a default so the first iteration will always run faster, but make it a hyperparameter so successive iterations can attempt to use it.

- building a heuristic for ARIMA that determines if the dimensionality/length of the dataset is below a certain point at which the fit time for ARIMA isn't drastically increased (<60%?) if covariates are used. If it is above that point, covariate usage will be set to False.