Tuning slow for huge data sets because backtest splits are created for each round

[janrth]

Description

For data sets with huge amounts of data, the tuning can be slow, because the backtest is created for each tuning round (as far as I see it) in mlforecast_objective.

Th core fitting methods are often fast because they are based on fast implementations from lgbm or xgb for example. But the backtest split can be a bottleneck.

I propose to leave the backtest split function as it is, but to re-use created splits (somehow) so the splits don't have to be created multiple times (as suggested by @yherin)

Use case

Faster tuning process due to reduction of multiple calculations of backtest splits.