Read all hdf5 event data on reader init.

Reading only single events as done before this commit is extremely slow,
i.e. many orders of magnitude slower. The downside is that we need quite
some memory to hold all events.