Backports for v0.14.4 (#3121)

*Description of changes:* backporting fixes
- #3094 
- #3115 

Additionally: cap pandas version to <2.2


By submitting this pull request, I confirm that you can use, modify,
copy, and redistribute this contribution, under the terms of your
choice.


**Please tag this pr with at least one of these labels to make our
release process faster:** BREAKING, new feature, bug fix, other change,
dev setup

---------

Co-authored-by: Jirka Borovec <[email protected]>

requirements/requirements-test.txt

@@ -1,4 +1,4 @@
-pandas>=1.1
+pandas>=1.1,<2.2.0
 flaky~=3.6
 pytest-cov==2.6.*
 pytest-timeout~=1.3

requirements/requirements.txt

@@ -1,5 +1,5 @@
 numpy~=1.16
-pandas>=1.0,<3
+pandas>=1.0,<2.2.0
 pydantic>=1.7,<3
 tqdm~=4.23
 toolz~=0.10

src/gluonts/time_feature/lag.py

@@ -146,7 +146,7 @@ def _make_lags_for_month(multiple, num_cycles=3):
             + _make_lags_for_hour(offset.n / (60 * 60))
         )
     else:
-        raise Exception("invalid frequency")
+        raise ValueError(f"invalid frequency | `freq_str={freq_str}` -> `offset_name={offset_name}`")
 
     # flatten lags list and filter
     lags = [

src/gluonts/torch/model/predictor.py

@@ -50,7 +50,7 @@ def __init__(
         forecast_generator: ForecastGenerator = SampleForecastGenerator(),
         output_transform: Optional[OutputTransform] = None,
         lead_time: int = 0,
-        device: str = "auto",
+        device: Union[str, torch.device] = "auto",
     ) -> None:
         super().__init__(prediction_length, lead_time=lead_time)
         self.input_names = input_names