t.pytorch.predict: various issues with reference data and model config

[ninsbl]

Investigate:

• [ ] why does positive offset result in maps not found (likely due to extraction of start and end times (start time is for some reason end time of the first input)
• [x] make sure that the order of reference data is applied as expected (negative offset means reference maps represent earlier points in time)
• [x] how to handle binary clasification (rounding (0.5 threshold) may be too restrictive, implement support for other threshold? Use scaling?) esp. with inverted classes (0: presence 1: absence)

[ninsbl]

See: #61 adds argmax transformation for binary clasification results (used by NORCE algorithm).

[ninsbl]

BTW. When a negative reference is used it makes sense that the _starttime of the resulting map is the _endtime of the previous (e.g. offset=-1) granule as that is the time window covered...