implement `train_test_split='fair-sliding'/'fair-all'`

[aaronspring]

Is your feature request related to a problem? Please describe. New paper Risbey et al. 2021 demonstrates how forecaster choices for time selection in bias reduction impacts skill.

Reference: https://t.co/HE4PoROb8Z?amp=1

Describe the solution you'd like hindcast.remove_bias(how='additive_mean', alignment='same_verifs', method='fair_sliding')

implemented method from:

• [x] unfair #655
• [x] unfair-cv #655
• [x] fair #655
• [ ] fair-sliding
• [ ] fair-all

unfair-cv is often standard, see #643, but unfair wrt real-time forecast/attainable skill, i.e. unfair means artificial skill

Demo: https://gist.github.com/aaronspring/94e24fb7e0a536be61caf1970287f3b5

[aaronspring]

API: method

• fair: train_init=slice(start, end) ~~,test_init=slice(start, end)~~ (need to change maybe to train_time for same_verifs)
• fair-sliding: train_init=slice(start, end),~~test_init=slice(start, end)~~
• fair-all: test_init=slice(start, end) ~~,train_init=slice(start, end)~~
• unfair-cv, cv='loo'
• unfair: now default