Better handling of small time series groups

[paxcema]

• Mindsdb version: current stable 2.41.0

For group-bys that have a high amount of combinations (e.g. origin-destination airports) it might be that the available data is less than the needed window to train the predictor. We should handle these cases so that the predictor can still train the groups that do have enough data (currently, the "not enough historical context" error interrupts training).

The groups that have insufficient information might still revert to using the default normalizer down the line, hopefully with decent enough forecasts with the current approach, though this would have to be confirmed.