UniTS model hard-coded "cuda:0" when loading pretrain weights

[zhangzw16]

ProbTS/probts/model/forecaster/point_forecaster/units.py

Line 1014 in d299e1c

ckpt = torch.load(pretrain_weight_path, map_location='cuda:0')

This should be made more robust to support various devices.

[zhangzw16]

Fixed with #24

[zhangzw16]

Maybe need a better way to specify device when inference.