Keep data lazy in class STDataset

[SarahAlidoost]

Currently we read data using xr.open_mfdataset and create dataset using STDataset. We can specify chunks in xr.open_mfdataset when data is large, but STDataset doesnot respect it. .values and from_numpy calls in STDataset class, data is not lazy and the benefit of chunking is mostly lost. We want refactor STDataset so data stays lazy, see this comment.

[SarahAlidoost]

if data is lazy STDataset class,, because of dask graph, the batch loop takes long and memory issue is still there. Si, instead of loading large data set to STDataset class, we loop over year because one year of data fits in memory on HPC.