adding jupyter notebook for dataset API, WIP

[svensht2]

//cc @limorhe @giadang

this is the direction I'm thinking for loading our datasets

[giadang]

I'm fine with this except the open/close the h5 files every time you call the function _read_hdf5_func, which may cause the overhead as people are talking about. Is there a way you can pre-open those h5 files and store them into something like a dictionary, and inside the function _read_hdf5_func you only need to get the right h5 object instead of opening/closing it again and again?