AEP 08 - New Forecasting module

[TonyBagnall]

AEP Template

Contributors: @TonyBagnall, Leo, Adam, Alex

Overview

We have recently removed the whole legacy forecasting module. We wish to start again
with a more tightly focussed goal in forecasting. Our priority is to develop a
simple model for implementing forecasters from stats, machine learning and deep
learning in a consistent way. We want to start simple and build things up, so we
start with the following assumptions.

1. We will begin with a focus on forecasting with a single series.
2. We will use numpy input only.
3. We focus on forecasting for longer series with longer windows.
4. We look for a clean train/test devision in fit/predict and will mostly avoid
   recursive models: forecasters trained on a given horizon are meant to forecast on
   that horizon

Problem Statement and Use Cases

the problem is to provide a simple interface for forecasting. This involves fitting
a model on some training data, then making forecasts.

There are three classes of algorithm for forecasting: statistical, machine learning
and deep learning. The main difference is that machine learning and deep learning
algorithms are primarily a reduction to regression through a sliding window.

Some notes on other implementations here
https://hackmd.io/@aeon-toolkit/rJZFioiJJe

more details in the PR