Fix edge case in PyTorchPredictor.deserialize

[lostella]

Description of changes: Since #2965 (not sure before), if a PyTorchPredictor object was serialized from the CPU memory, then deserializeing it with device="cuda" would actually not work. This would happen:

1. predictor object created, with parameters of its torch.nn.Module model allocated on CPU (since that was the device the predictor was serialized with)
2. when deserializing with device="cuda", torch.load with map_location="cuda" would put the state_dict values on GPU as expected
3. torch.nn.Module.load_state_dict would however copy the parameters back to CPU
4. since the predictor has the .device attribute set to "cpu", prediction would not complain (data & model on the same device) but would be really slow.

This PR makes sure that the predictor object being created is moved .to(device) after step 1, so that step 3 actually keeps parameters on GPU.

The same issue happens inverting CPU and GPU, as in the following example

import pandas as pd
import numpy as np
import logging
import tempfile
from gluonts.model import Predictor
from gluonts.torch.model.wavenet import WaveNetEstimator
from pathlib import Path

logging.basicConfig(level=logging.INFO)

data = [
    {
        "start": pd.Period("2012-02-04", freq="D"),
        "target": np.ones(2000),
    }
]

estimator = WaveNetEstimator(
    freq="H",
    prediction_length=24,
    num_batches_per_epoch=2,
    trainer_kwargs=dict(max_epochs=1),
)

predictor = estimator.train(data)

with tempfile.TemporaryDirectory() as td:
    predictor.serialize(Path(td))
    predictor = Predictor.deserialize(Path(td), device="cpu")
    print(f" predictor device is {predictor.device}")
    print(f"module parameters on {next(predictor.prediction_net.parameters()).device}")

Expected to have the resulting model on CPU. Output before the PR:

 predictor device is cuda
module parameters on cuda:0

Output after the PR:

 predictor device is cpu
module parameters on cpu

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

[lostella]

@abdulfatir now I'm wondering, whether it makes sense to have a device option in PyTorchPredictor.deserialize at all, since doing

predictor = Predictor.deserialize(path, device=device)

is the same as doing

predictor = Predictor.deserialize(path)
predictor.to(device)