refactored preprocessor/encoder/decoder and fixed some bugs

[LeungTsang]

preprocessor:

1. preprocessor is separated from dataset. dataset is no longer recursive. instead, dataset accepts a preprocessor object for data preprocessing
2. transformation can be defined in arbitary order. e.g. crop and band filter can be the first step to reduce computation.
3. tile is replaced by sliding inference in evalutor, which avoids loading images from disk multiple times and allows generating complete prediction for a whole image.
4. enable verification of data size and dimension

encoder/decoder:

1. incorporate multi temporal forward in encoder implementation instead
2. squeeze temporal dimension safely

fix some bugs:
multi temporal output
save best model

TO DO:
check if everything works
merge segmentation and regression tasks as much as possible
fix bugs e.g. regevaluator