Fix get_index_root_data() to work with out-of-order trains in files

In files, trains are not guaranteed to be in order - while all public EXtra-data APIs compensate for this, the FileAccess API presents a low-level view of a single file and thus still includes it.

The call to np.intersect1d used to determine the subset of trains in each sequence file always returns a sorted result for trains, and thus an unsorted index array. This breaks for direct access to h5py datasets.

This MR transfers that order back to the index array from the subset of trains in a sequence file as suggested by @danilo

@danilo

This now passes for the command line posted in !15 (comment 453642)

LGTM

merged

mentioned in commit b6bfd406