[Models]: provide an (non-auto)ARIMA model class as well

[almostintuitive]

Description

We'd find it useful if we could:

1. re-use the parameters found by AutoARIMA
2. sometimes, simply provide the parameters ourselves

and use your efficient implementation, instead of statsmodel's. We think this could be achieved by providing a simple ARIMA class that we can use to manual input parameters to the ARIMA model (we do understand that there's a function we could use, but it looks like the class interface is the canonical way to access it).

Thank you!

Use case

No response

[canbooo]

Came here to raise this exact issue. Here is a use case: I would like to obtain the results of the auto search (i.e. model_["arma"]) once using a training set but keep these results constant during cross-validation with a validation/dev set since the search takes quite long time with longer time series. Although the effect is not as large, similar arguments could be made about the other auto-models. Alternatively, allow setting p, q, d, P, Q, D to constant values or allow partial fit of the model without the search.