`select_trains()` is slow

In load or open_run, we use run.select_trains(subset) with subset=ed.by_index[:] by default. This is not necessary and takes a substantial amount of time:

For a run with 72000 trains, open_run takes 13 seconds, whil extra_data open_run takes 126 ms. select_trains should only be used when not all trains are wanted.