To be sure can we use location as feature for training and inference, or create different model location wise? Is it acceptable?
Yes, you can create separate models for different locations or use location as a feature.
Similarly can we use grid_ids or generic encoding of grid_id derived from data as features?
Yes, but be cautioned that there are grid IDs present in the NO2 test set that are not present in the NO2 train set. Any solution should be able to adapt to missing features, including grid ids.