t.pytorch.predict: Add options for raster_maps and reference_strds as input

[ninsbl]

In order to be able to use reference images in predictions (e.g. for flood or avalances), options for _rastermaps and _referencestrds (the latter together with a _spatio-temporalrelation) should be added.

• _rastermaps: Here a comma-separted list of raster maps could be provided by the user. Those _rastermaps will then be added to every image group and matched to model input using semantic_labels (examples for data provided as _rastermaps could be elevation, slope, land_cover...).
• _referencestrds: Reference images may be time-dependent. So, the idea is, if a _referencestrds is given, together with a _spatio-temporalrelation between the reference and the input STDRS, the matching reference raster maps should be added reference imagesto every image group and matched to model input using semantic_labels (examples for data provided as _referencestrds could be backscatter intensity of the previous repeat-pass for Sentinel-1, average backscatter of the previous year, ...)..

[ninsbl]

We also need a good solution to handle semantic labels. Sentinel-1 data have semantic labels with polarization, trac, and direction information (e.g. "VV_151_descending"). So, semantic labels should match between "activity" and "reference", and relation between "activity" and "reference" (see: G_OPT_T_SAMPLE, equal,overlaps,overlapped,starts,started,finishes,finished,during,contains)