Missing value handling in `arx_forecaster()`?

[ryantibs]

Should we implement an option for missing value handling:

• at training time in arx_forecaster()?
• at prediction time in arx_forecaster()?

In the latter, for example, we should for example be careful that we don't allow LOCF to be done over arbitrarily long time stretches. For example, there should allow the user to set a max amount of time that they are comfortable to use for propagating an old value forward; beyond that time, we get an NA, and the whole prediction is NA.