Backporting fixes:
- Fix: avoid automatic device detection via serialized tensors when deserializing PyTorchPredictor. #2576 by @shubhamkapoor
- Fix Map representation. #2579 by @jaheba
- serde: Fix encoding of dtypes. #2586 by @jaheba
- Add assertion to split function ensuring valid windows #2587 by @MarcelK1102
- Ensure dtype on feat_time in torch DeepAR. #2596 by @jaheba
- Expose aggregation method in ensemble NBEATS, fix forecast shape #2598 by @lostella
- Fix requirements following breaking change in setuptools #2604 by @lostella
- Add gluonts.util.safe_extract #2606 by @lostella