update DataSet and DataSetLoader

[choosewhatulike]

now we can use different DataSetLoader to load data from files and convert into DataSet

[FengZiYjun]

For sequence labeling, in training, set_target(truth=False). In testing, set_target(truth=True).
This is because sequence labeling model with CRF computes loss in its forward() if truth given, and the loss needs an intermediate result.
But the naming is quite wired.
Can we develop a better machenism to handle this?

Currently,

If only x is given, output prediction.
x ----(forward)-----(viterbi)---> prediction

If x and ground truth are given, output loss.
x -----(forward)----(crf)-->  loss
truth

Before,

                                      truth
x ----(forward)----> y  ---(crf)----> y_t ------> loss
                    |---(viterbi)---> prediction

Maybe a tricky way,

x ---- (forward) ----->  y (record this value) -----(viterbi, still in forward)----> prediction
y + truth ----(crf, in loss method)---->  loss

[FengZiYjun]

agree to merge